Power transmission line fault detection method and system based on RKM-YOLO algorithm

Abstract

The present application relates to the technical field of image processing, and particularly relates to a power transmission line fault detection method and system based on an RKM-YOLO algorithm. An intelligent agent for power transmission line fault detection based on an RKM-YOLO algorithm is acquired. A power transmission line image collected by current power transmission line inspection is acquired, the power transmission line image is input into the intelligent agent for power transmission line fault detection, and a fault mode is obtained by recognition, and local feature scale extraction and overall feature scale extraction are performed on the power transmission line image. Reference local feature scales and reference overall feature scales corresponding to the fault mode recognized in a historical time are acquired, similarity with the reference local feature scales and the reference overall feature scales is analyzed, and is used as a detection confidence, and the local feature scales and the overall feature scales are self-checked, a correction confidence is obtained by correcting the detection confidence, and confidence labeling is performed. The present application can realize high-precision and high-efficiency detection of power transmission line faults.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and in particular to a method and system for detecting transmission line faults based on the RKM-YOLO algorithm. Background Technology

[0002] Traditional transmission line fault detection relies heavily on UAV inspection images and conventional target detection models for fault identification. This method suffers from high rates of missed detection and false detection of small targets in complex environments, as well as high computational complexity, making it difficult to meet the requirements for efficient and accurate fault detection and impacting the efficiency of power grid safety operation and maintenance. Summary of the Invention

[0003] This invention addresses the problem that existing technologies struggle to meet the demands for efficient and accurate fault detection by providing a transmission line fault detection method and system based on the RKM-YOLO algorithm.

[0004] The technical solution of the present invention to solve the above-mentioned technical problems is as follows:

[0005] In a first aspect, the present invention provides a transmission line fault detection method based on the RKM-YOLO algorithm, comprising: acquiring a transmission line fault detection agent constructed based on the RKM-YOLO algorithm; acquiring transmission line images collected during current transmission line inspections, inputting the transmission line images into the transmission line fault detection agent, outputting a fault identification mode, and extracting local and global feature scales from the transmission line images; acquiring a baseline local feature scale and a baseline global feature scale corresponding to the fault identification mode within a historical time period, analyzing the similarity with the baseline local and global feature scales as a detection confidence level, performing self-verification on the local and global feature scales, correcting the detection confidence level to obtain a corrected confidence level, labeling the confidence level, and completing the fault detection.

[0006] Optionally, obtaining a transmission line fault detection agent based on the RKM-YOLO algorithm includes: constructing the main architecture of the transmission line fault detection agent based on the YOLOv8 model; introducing the RFCAConv module within the main architecture to first perform group convolution on the input features to extract receptive field features; replacing the original C2f module in YOLOv8 with the K3_RFCA module, which adopts a dual parallel 1×1 convolution branch structure and integrates the RFCAConv module, fusing shallow and deep features through residual connections; introducing a multi-scale multi-dimensional collaborative attention module (MMCA) within the main architecture, using multi-scale convolution and spatial-channel interaction mechanisms to extract multi-scale features through depthwise separable convolution, and employing an adaptive weighting strategy to strengthen the representation of key regions; and obtaining the transmission line fault detection agent through supervised training based on the main architecture, the RFCAConv module, and the multi-scale multi-dimensional collaborative attention module (MMCA).

[0007] Optionally, the process involves acquiring images of the transmission line currently being inspected, inputting these images into the transmission line fault detection agent, outputting a fault identification mode, and extracting local and global feature scales from the images. This includes: acquiring images of the transmission line currently being inspected; inputting these images into the transmission line fault detection agent, outputting a fault identification mode; acquiring a transmission line scale extractor, wherein the scale extractor includes a local scale extraction branch and a global scale extraction branch; inputting the images into the local scale extraction branch and the global scale extraction branch, outputting local and global feature scales, wherein the local feature scale is the average scale of multiple sheds on the insulator, and the global feature scale is the scale of the insulator itself.

[0008] Optionally, obtaining a transmission line scale extractor includes: collecting a set of sample transmission line images based on historical inspection data of transmission lines; labeling the local and global feature scales within each sample transmission line image to obtain a set of sample local feature scales and a set of sample global feature scales; constructing a local scale extraction branch and a global scale extraction branch based on a convolutional neural network; using the set of sample transmission line images as input and the set of sample local feature scales and the set of sample global feature scales as outputs, respectively, supervising the training of the local scale extraction branch and the global scale extraction branch, and obtaining the transmission line scale extractor after convergence.

[0009] Optionally, obtaining the baseline local feature scale and baseline global feature scale corresponding to the identified fault mode within a historical time period includes: obtaining all historical transmission line images within a historical time period that have been correctly identified and verified by the transmission line fault detection agent; extracting the historical local feature scale set and historical global feature scale set of all historical transmission line images; calculating the occurrence rate of each type of historical local feature scale and historical global feature scale in the historical local feature scale set and historical global feature scale set, and selecting the historical local feature scale and historical global feature scale with the highest occurrence rate as the baseline local feature scale and baseline global feature scale.

[0010] Optionally, the similarity between the local feature scale and the benchmark global feature scale is analyzed as the detection confidence score, including: calculating the similarity between the local feature scale and the benchmark local feature scale as the local similarity; calculating the similarity between the global feature scale and the benchmark global feature scale as the global similarity; and fusing the local similarity and the global similarity to obtain the detection confidence score.

[0011] Optionally, self-verification is performed on the local feature scale and the global feature scale to correct the detection confidence and obtain a corrected confidence, including: calculating the estimated global feature scale based on the local feature scale; calculating the estimated local feature scale based on the global feature scale; and performing self-verification based on the local feature scale, the global feature scale, the estimated global feature scale, and the estimated local feature scale to correct the detection confidence and obtain a corrected confidence.

[0012] Specifically, based on the local feature scale, the overall feature scale, the inferred overall feature scale, and the inferred local feature scale, self-verification is performed to correct the detection confidence and obtain a corrected confidence. This includes: calculating the similarity between the local feature scale and the inferred local feature scale as a local verification score; calculating the similarity between the overall feature scale and the inferred overall feature scale as an overall verification score; fusing the local verification score and the overall verification score to obtain a verification score, and calculating a correction coefficient; and using the correction coefficient to correct the detection confidence to obtain a corrected confidence.

[0013] Optionally, confidence level labeling can be performed to complete fault detection, including:

[0014] The identified fault patterns are labeled using the corrected confidence level to complete fault detection.

[0015] Secondly, this invention provides a transmission line fault detection system based on the RKM-YOLO algorithm, comprising:

[0016] The agent acquisition module is used to acquire a transmission line fault detection agent constructed based on the RKM-YOLO algorithm;

[0017] The fault mode recognition module is used to acquire images of the transmission line currently being inspected, input the images of the transmission line into the transmission line fault detection agent, output the identified fault mode, and perform local feature scale extraction and global feature scale extraction on the images of the transmission line.

[0018] The confidence labeling module is used to obtain the baseline local feature scale and baseline global feature scale corresponding to the identified fault mode within a historical time period, analyze the similarity with the baseline local feature scale and baseline global feature scale as the detection confidence, perform self-verification on the local feature scale and global feature scale, correct the detection confidence to obtain the corrected confidence, perform confidence labeling, and complete the fault detection.

[0019] By implementing this invention, it is possible to obtain a transmission line fault detection intelligent agent built based on the RKM-YOLO algorithm. This intelligent agent enhances the feature extraction capability in complex environments by leveraging various modules, improves the sensitivity of small targets by integrating deep and shallow features, and strengthens the representation of key areas, laying a solid foundation for subsequent accurate fault identification and providing an efficient and reliable core tool for fault detection.

[0020] By implementing this invention, it is possible to acquire images of transmission lines currently being inspected, input these images into a transmission line fault detection intelligent agent, output the identified fault modes, and perform local and global feature scale extraction on the images. This not only enables rapid identification of potential fault types in the current images but also provides crucial data support for subsequent comparison with benchmark scales and assessment of detection confidence by extracting the average scale of insulator shrouds and the scale of the insulators themselves, making fault detection more data-driven.

[0021] By implementing this invention, it is possible to obtain the baseline local feature scale and baseline global feature scale corresponding to the identified fault mode within a historical time period, analyze the similarity with the baseline local feature scale and baseline global feature scale as the detection confidence level, perform self-verification on the local feature scale and global feature scale, correct the detection confidence level to obtain the corrected confidence level, perform confidence level labeling, complete the fault detection, combine with historical valid data to ensure that the confidence level judgment has a reference, further correct errors using self-verification to improve the accuracy of the confidence level, and finally label the confidence level to make the fault detection results more reliable, thereby reducing misjudgment and missed judgment, and ensuring the quality of fault detection.

[0022] In summary, by implementing this invention, high-precision and high-efficiency detection of transmission line faults can be achieved, providing intelligent support for the safe operation and maintenance of the power grid. Attached Figure Description

[0023] Figure 1 This is a flowchart illustrating the transmission line fault detection method based on the RKM-YOLO algorithm provided by the present invention.

[0024] Figure 2 This is a schematic diagram of the structure of the transmission line fault detection system based on the RKM-YOLO algorithm provided by the present invention.

[0025] In the attached diagram, the components represented by each number are as follows:

[0026] The system includes an agent acquisition module 11, a fault mode recognition module 12, and a confidence labeling module 13. Detailed Implementation

[0027] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0028] In the description of this invention, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more features. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0029] In the description of this invention, the term "for example" is used to mean "used as an example, illustration, or description." Any embodiment described as "for example" in this invention is not necessarily to be construed as being more preferred or advantageous than other embodiments. The following description is provided to enable any person skilled in the art to make and use the invention. Details are set forth in the following description for purposes of explanation. It should be understood that those skilled in the art will recognize that the invention can be made without using these specific details. In other instances, well-known structures and processes will not be described in detail to avoid obscuring the description of the invention with unnecessary detail. Therefore, the invention is not intended to be limited to the embodiments shown, but is consistent with the broadest scope of the principles and features disclosed herein.

[0030] Example 1, as Figure 1As shown, this embodiment of the invention provides a transmission line fault detection method based on the RKM-YOLO algorithm, including:

[0031] S100: Obtain the intelligent agent for transmission line fault detection constructed based on the RKM-YOLO algorithm;

[0032] S200: Acquire the transmission line image collected during the current transmission line inspection, input the transmission line image into the transmission line fault detection agent, output the identified fault mode, and perform local feature scale extraction and global feature scale extraction on the transmission line image;

[0033] S300: Obtain the baseline local feature scale and baseline global feature scale corresponding to the identified fault mode within a historical time period, analyze the similarity with the baseline local feature scale and baseline global feature scale as the detection confidence, perform self-verification on the local feature scale and global feature scale, correct the detection confidence to obtain the corrected confidence, perform confidence labeling, and complete the fault detection.

[0034] In step S100 of this application embodiment, obtaining a transmission line fault detection intelligent agent constructed based on the RKM-YOLO algorithm includes:

[0035] Based on the YOLOv8 model, the main architecture of the transmission line fault detection agent is constructed.

[0036] Within the main architecture, the RFCAConv module is introduced to first perform group convolution on the input features to extract receptive field features;

[0037] The original C2f module in YOLOv8 is replaced by the K3_RFCA module. The K3_RFCA module adopts a dual parallel 1×1 convolutional branch structure and integrates the RFCAConv module. Shallow and deep features are fused through residual connections.

[0038] The main architecture introduces a multi-scale multi-dimensional collaborative attention module (MMCA), a multi-scale convolution and spatial-channel interaction mechanism, extracts multi-scale features through depthwise separable convolution, and adopts an adaptive weighting strategy to enhance the representation of key regions.

[0039] Based on the aforementioned main architecture, RFCAConv module, and multi-scale multi-dimensional collaborative attention module MMCA, a transmission line fault detection agent is obtained through supervised training.

[0040] In this embodiment of the application, the purpose of step S100 is to construct a transmission line fault detection agent based on the RKM-YOLO algorithm. This agent needs to have the ability to accurately extract transmission line features in complex environments in order to improve the identification effect of faults in small target components, while reducing the computational complexity of the model, laying the foundation for accurate detection of transmission line faults in the future, and meeting the real-time processing requirements of UAV inspection.

[0041] To achieve the above objectives, the first step is to construct the main architecture of the transmission line fault detection agent based on the YOLOv8 model. This means directly using the YOLOv8 model as the basic framework, retaining its original network layer structure, detection process, and other core components, and defining this framework as the main architecture of the transmission line fault detection agent. Subsequent module improvements and optimizations will all be carried out on this main architecture.

[0042] The second step involves introducing the RFCAConv module within the main architecture to perform group convolutions on the input features to extract receptive field features.

[0043] First, group convolution operations are performed on the power transmission line image features input to the main architecture. This divides the feature map into multiple groups, each convolved separately to extract feature information from different receptive fields. Then, ReLU activation is applied to the receptive field features extracted by group convolution to enhance their non-linear expressive power. Simultaneously, a coordinate attention mechanism is introduced, embedding positional information into channel attention, enabling the model to focus on feature location and improving sensitivity to features in specific regions. Finally, a softmax function is used to weight each receptive field feature, highlighting the role of important features, thus completing the integration and feature processing of the RFCAConv module within the main architecture.

[0044] The third step is to replace the original C2f module in YOLOv8 with the K3_RFCA module. The K3_RFCA module adopts a dual parallel 1×1 convolutional branch structure, integrates the RFCAConv module, and fuses shallow and deep features through residual connections.

[0045] Specifically, a dual-parallel 1×1 convolutional branch structure is designed, in which the previously introduced RFCAConv module is integrated. Then, the original C2f module in the YOLOv8 model is removed and replaced with the designed K3_RFCA module. Within the K3_RFCA module, a residual connection mechanism is used to fuse shallow features with the deep features processed by the RFCAConv module, ensuring that the fused features simultaneously contain both shallow details and deep semantic information, thus completing the module replacement and feature fusion process.

[0046] The fourth step is to introduce a multi-scale, multi-dimensional collaborative attention module (MMCA) into the main architecture, which uses multi-scale convolution and spatial-channel interaction mechanism to extract multi-scale features through depthwise separable convolution and adopts an adaptive weighting strategy to enhance the representation of key regions.

[0047] Specifically, a multi-scale, multi-dimensional collaborative attention module (MMCA) is built within the main architecture. This module first uses depthwise separable 5×5 convolutions to process the input features, extract local features, and capture fine-grained information about the target, such as subtle defects in the insulator's shed disc.

[0048] Then, convolution operations of 1×7, 7×1, 1×11, 11×1, 1×21 and 21×1 are used to extract feature information at different scales, covering local details at small scales and overall structural features at large scales.

[0049] Simultaneously, a spatial-channel interaction mechanism is introduced, which rotates the spatial dimension of the input features by 90° to obtain the rotated feature map. Aggregated feature weights are generated through channel compression and excitation transformation, and then the sigmoid function is applied to convert these weights into attention weights.

[0050] Then, a dynamic pooling strategy combining global average pooling and standard deviation pooling is adopted to dynamically aggregate multi-scale contextual information. At the same time, an adaptive weighting strategy is used to assign different weights to different features to strengthen the representation of key regions, thus completing the introduction and feature processing of the MMCA module.

[0051] The fifth step involves supervising the training of the transmission line fault detection agent based on the main architecture, the RFCAConv module, and the multi-scale multi-dimensional collaborative attention module MMCA.

[0052] This involves constructing a sample library of transmission line components and typical faults. The sample library contains a large number of transmission line image samples with different scenarios and fault types, such as insulator breakage, pollution flashover, and vibration damper corrosion. Information such as fault type and feature scale in the transmission line image samples is labeled.

[0053] The labeled transmission line image samples are input into a model that integrates the main architecture, RFCAConv module and MMCA module. The difference between the fault identification result output by the model and the sample labeling information is used as the loss function. The parameters in the model, such as the weights and learnable parameters in each module, are continuously adjusted through backpropagation.

[0054] The training process continues iteratively until the model's loss function value converges to a preset level, and the performance indicators such as fault detection accuracy and recall on the validation set reach preset standards. At this point, training is stopped, and the intelligent agent for detecting power transmission line faults is obtained.

[0055] In step S200 of this application embodiment, the transmission line image collected during the current transmission line inspection is obtained, the transmission line image is input into the transmission line fault detection agent, the identified fault mode is output, and local feature scale extraction and global feature scale extraction are performed on the transmission line image, including:

[0056] Acquire images of the transmission lines currently being inspected;

[0057] The image of the power transmission line is input into the power transmission line fault detection intelligent agent, and the identified fault mode is output.

[0058] A transmission line scale extractor is obtained, wherein the transmission line scale extractor includes a local scale extraction branch and a global scale extraction branch;

[0059] The image of the transmission line is input into the local scale extraction branch and the global scale extraction branch, and the local feature scale and the global feature scale are output. The local feature scale is the average scale of multiple sheds on the insulator, and the global feature scale is the scale of the insulator.

[0060] In this embodiment of the application, the purpose of step S200 is to acquire images of the transmission line in the current inspection scenario, preliminarily identify the fault type with the help of the constructed fault detection intelligent agent, and extract the local and global feature scales of transmission line components such as insulators in the image through a special scale extractor, so as to lay a data foundation for subsequent verification of the accuracy of fault identification results by feature scale comparison and to improve the fault detection accuracy in complex environments.

[0061] To achieve the above steps, it is first necessary to acquire images of the power transmission lines currently being inspected. Specifically, this can be done by using a drone equipped with image acquisition equipment to perform the current power transmission line inspection task. Following a preset inspection path and with a preset shooting resolution and frame rate, the drone can acquire images of components along the power transmission line, such as insulators and vibration dampers. Once the acquisition is complete, these real-time generated images of the power transmission lines can be directly obtained.

[0062] Then, the transmission line image needs to be input into the transmission line fault detection agent, and the identified fault mode needs to be output. The acquired current transmission line image is preprocessed according to the agent's requirements, such as size normalization and pixel value standardization, and then input into the transmission line fault detection agent built based on the RKM-YOLO algorithm. The agent uses its internal RFCAConv module to extract anti-interference features, the K3_RFCA module to fuse deep and shallow features, and the MMCA module to enhance the representation of key regions to identify and classify fault features in the image, finally outputting the corresponding identified fault mode, such as no fault, insulator fracture fault, vibration damper corrosion fault, etc.

[0063] Next, a transmission line scale extractor needs to be obtained.

[0064] In step S200 of this application embodiment, obtaining the transmission line scale extractor includes:

[0065] Based on historical inspection data of transmission lines, a set of sample transmission line images is collected. The local feature scale and global feature scale of each sample transmission line image are labeled to obtain the sample local feature scale set and the sample global feature scale set.

[0066] Based on convolutional neural networks, local scale extraction branches and global scale extraction branches are constructed.

[0067] Using the sample transmission line image set as input, and the sample local feature scale set and sample global feature scale set as output, the local scale extraction branch and the global scale extraction branch are trained under supervision, and a transmission line scale extractor is obtained after convergence.

[0068] In this embodiment, the transmission line scale extractor is a tool that can accurately extract the local and overall feature scales of transmission line components. It solves the problem that changes in drone shooting angle and lighting can cause component scale deformation, which in turn affects the accuracy of fault identification. It provides stable and accurate scale extraction capabilities to support subsequent verification of the reliability of fault identification results through feature scale comparison.

[0069] To obtain the transmission line scale extractor, it is first necessary to collect a set of sample transmission line images based on historical inspection data of transmission lines, and then label the local feature scale and the overall feature scale within each sample transmission line image to obtain the sample local feature scale set and the sample overall feature scale set.

[0070] This involves selecting images of transmission lines from a historical inspection database, covering different seasons, weather conditions, shooting angles, fault types, and normal states, to form a sample transmission line image set. For each sample image, using image annotation tools, local feature scales of the target components are manually or semi-automatically annotated, such as the individual scales of multiple sheds of an insulator, as well as overall feature scales, such as the overall length and width of the insulator. The local feature scales of all samples are then compiled into a sample local feature scale set, and the overall feature scales are compiled into a sample overall feature scale set.

[0071] Then, based on the convolutional neural network, it is necessary to construct local scale extraction branches and global scale extraction branches.

[0072] Optionally, the local scale extraction branch includes the following structure:

[0073] The input layer is used to receive images of power transmission lines. Convolutional layer 1 has 32 kernels, a kernel size of 3×3, a stride of 1, and uses ReLU activation function to initially extract shallow local features of the image.

[0074] Max pooling layer 1, with a pooling kernel size of 2×2 and a stride of 2, serves to reduce the dimension of the feature map, reduce computational cost, and preserve local feature information.

[0075] Convolutional layer 2, with 64 convolutional kernels, a kernel size of 3×3, a stride of 1, and the activation function ReLU, further extracts more refined local features, such as the edge details of the insulator shed.

[0076] Max pooling layer 2, with a pooling kernel size of 2×2 and a stride of 2. Convolutional layer 3, with 128 convolutional kernels, a kernel size of 3×3, and a stride of 1. The activation function used is ReLU to enhance the capture of local features of the insulator's umbrella disk. Max pooling layer 3, with a pooling kernel size of 2×2 and a stride of 2.

[0077] Four convolutional layers are used, with 256 kernels of size 3×3 and a stride of 1. The ReLU activation function is used for depthwise extraction of local features, providing support for subsequent scale calculations. A global average pooling layer is used to aggregate local feature information and reduce the number of parameters.

[0078] Fully connected layer 1 has 128 neurons and uses the ReLU activation function to perform a non-linear transformation on the aggregated feature vector.

[0079] The fully connected layer 2 has 1 neuron, no activation function, and outputs the predicted value of local feature scale, which is the average scale of multiple umbrella disks on the insulator.

[0080] Optionally, the overall scale extraction branch includes the following structure: an input layer that receives the transmission line image; a single convolutional layer with 32 kernels (7×7 kernel size) and a stride of 1, using ReLU activation to initially extract shallow overall features of the image; and a single max-pooling layer with a 2×2 kernel size and a stride of 2.

[0081] Convolutional layer 2, with 64 convolutional kernels of size 5×5 to maintain a large receptive field, a stride of 1, and ReLU activation function, is used to further extract the overall structural features of the insulator. Max pooling layer 2, with a pooling kernel size of 2×2 and a stride of 2.

[0082] Convolutional layer 3 has 128 kernels, each with a size of 5×5 and a stride of 1. The activation function is ReLU, which is used to enhance the feature capture of the overall contour of the insulator.

[0083] The maximum pooling layer is 3, where the pooling kernel size is 2×2 and the step size is 2.

[0084] Convolutional layer 4, with 256 convolutional kernels, a kernel size of 5×5, a stride of 1, and the activation function ReLU, is used for depth extraction of overall features, providing support for overall scale calculation.

[0085] The global average pooling layer is used to convert the 80×80×256 feature map into a 1×1×256 feature vector, aggregating the overall feature information.

[0086] Fully connected layer 1, with 128 neurons and ReLU activation function, performs a non-linear transformation on the aggregated feature vector.

[0087] The fully connected layer 2 has 1 neuron and no activation function. Its output is the overall feature scale prediction value, i.e., the scale of the insulator.

[0088] For the training of the local scale extraction branch and the global scale extraction branch, the Adam optimizer was used with an initial learning rate of 0.001 and a cosine annealing decay strategy. The learning rate was adjusted after each training round based on the training progress to avoid overfitting in later stages. The β1 parameter was 0.9, the β2 parameter was 0.999, and the weight decay coefficient was 0.0005. The mean squared error loss function was used because it effectively measures the magnitude of the error between the predicted scale and the true labeled scale, making it suitable for scale regression tasks. The batch size was set to 16 to balance training efficiency and gradient stability while ensuring sufficient computational resources.

[0089] At least 10,000 sample transmission line images were collected to form a sample transmission line image set, with a training set to validation set ratio of 4:1. Each image was labeled with its corresponding local and global feature scales. The initial training epochs were set to 100. If the model converged early during training, training could be stopped early.

[0090] When the loss function value on the validation set remains stable for 10 consecutive rounds, and the average absolute error between the predicted local feature scale value and the true value on the validation set is less than 5 pixels, and the average absolute error between the predicted overall feature scale value and the true value is less than 8 pixels, the model is considered to have converged, and the transmission line scale extractor is obtained.

[0091] Furthermore, the transmission line image needs to be input into the local scale extraction branch and the global scale extraction branch, and the outputs are the local feature scale and the global feature scale. The local feature scale is the average scale of multiple sheds on the insulator, and the global feature scale is the scale of the insulator.

[0092] The current transmission line image is input into the local scale extraction branch and the global scale extraction branch of the transmission line scale extractor. The local scale extraction branch extracts the local features of the insulator sheds in the image step by step through convolutional layers, calculates the scale of multiple sheds and takes the average value to obtain the local feature scale; the global scale extraction branch extracts the overall structural features of the insulator through convolutional layers, and calculates the overall feature scale of the insulator by combining the mapping relationship between image pixels and actual size; the two branches output the corresponding local feature scale and global feature scale, respectively.

[0093] Suppose a drone captures an image of an insulator on a power transmission line during an inspection. This insulator is a common disc insulator with 12 umbel discs. The umbel discs are disc-shaped structures on the insulator used to increase the creepage distance. Feature scales need to be obtained through local scale extraction and global scale extraction branches, respectively. The specific process is as follows:

[0094] After inputting the insulator image into the local scale extraction branch, the branch first locates the positions of all 12 umbrella discs in the image, and then calculates the key scales of each umbrella disc, such as the diameter of the umbrella disc. Assuming that the measured diameters of the umbrella discs are 150mm, 152mm, 149mm, 151mm, 150mm, 148mm, 153mm, 151mm, 149mm, 152mm, 150mm, and 149mm respectively, the average value of these 12 umbrella disc diameters is then taken, i.e., [150+152+149+151+150+148+153+151+149+152+150+149]÷12=1804÷12≈150.3mm. This value is the local feature scale.

[0095] After inputting the same insulator image into the overall scale extraction branch, the branch will locate the overall outline of the insulator, that is, the complete structure from the top connection end to the bottom connection end of the insulator, and calculate its key overall scale, such as the total length of the insulator. Assuming that the measured total length is 1800mm, this value is the overall feature scale.

[0096] In step S300 of this application embodiment, obtaining the baseline local feature scale and baseline global feature scale corresponding to the identified fault mode within a historical time period includes:

[0097] Acquire all historical images of transmission lines within a historical period that were obtained by the transmission line fault detection agent to identify the identified fault modes and verify their correctness.

[0098] Extract historical local feature scale sets and historical global feature scale sets from all historical transmission line images;

[0099] The occurrence rate of each type of historical local feature scale and historical overall feature scale in the historical local feature scale set and the historical overall feature scale set is statistically analyzed. The historical local feature scale and historical overall feature scale with the highest occurrence rate are selected as the benchmark local feature scale and benchmark overall feature scale.

[0100] In step S300 of this embodiment, the purpose of the above-mentioned detailed steps is to provide a reliable scale reference standard for assessing the confidence level of the current transmission line fault detection results. By determining the most representative local and global feature scales under a specific fault mode, namely the reference local feature scale and the reference global feature scale, these serve as reference scales. Subsequently, the reliability of the fault identification results can be judged by comparing the similarity between the currently detected feature scale and the reference scale, reducing misjudgments caused by scale deformation due to changes in drone shooting angle and lighting, and ensuring the accuracy of fault detection.

[0101] To achieve the above objectives, it is first necessary to acquire all historical transmission line images obtained by the transmission line fault detection agent within a historical period, which identify the fault patterns and verify their correctness.

[0102] This involves selecting historical transmission line images from the historical inspection database that meet two conditions: first, the fault detection agent outputs a result consistent with the currently identified fault mode; and second, the identification result has been verified as correct through manual review and professional equipment testing. All historical transmission line images meeting these two conditions are then compiled to form a valid image set for extracting the baseline scale.

[0103] Then, it is necessary to extract the historical local feature scale set and the historical global feature scale set of all historical transmission line images.

[0104] All the historical transmission line images compiled in the previous step are then input one by one into the trained transmission line scale extractor. The local scale extraction branch of the extractor calculates the local feature scale of target components, such as insulators, in each historical transmission line image, i.e., the average scale of multiple sheds. Then, the global scale extraction branch calculates the overall feature scale of the target components in each image, i.e., the overall scale of the insulator. Finally, the local feature scales of all historical images are compiled into a historical local feature scale set, and the global feature scales are compiled into a historical global feature scale set.

[0105] Next, it is necessary to calculate the occurrence rate of each type of historical local feature scale and historical overall feature scale in the historical local feature scale set and the historical overall feature scale set, and select the historical local feature scale and historical overall feature scale with the highest occurrence rate as the benchmark local feature scale and benchmark overall feature scale.

[0106] This involves classifying and statistically analyzing the historical local feature scale set. First, all local scale values ​​are categorized according to preset scale intervals, such as 0.1 mm intervals. Then, the occurrence rate is calculated as the proportion of the scale value appearing in each interval to the total occurrences. The same classification and statistical operation is performed on the historical overall feature scale set to obtain the occurrence rate of each category of overall feature scale. Typical values ​​corresponding to the scale intervals with the highest occurrence rates in the historical local feature scale set, such as the midpoint value of the interval, are selected as the baseline local feature scale. Similarly, typical values ​​corresponding to the scale intervals with the highest occurrence rates in the historical overall feature scale set are selected as the baseline overall feature scale.

[0107] In step S300 of this application embodiment, the similarity with the reference local feature scale and the reference global feature scale is analyzed as the detection confidence level, including:

[0108] Calculate the similarity between the local feature scale and the reference local feature scale, and use it as the local similarity.

[0109] Calculate the similarity between the overall feature scale and the benchmark overall feature scale, and use it as the overall similarity.

[0110] The local similarity and overall similarity are fused to obtain the detection confidence score.

[0111] In step S300 of this embodiment, the purpose of the above-mentioned subdivision step is to generate a detection confidence level that reflects the reliability of the fault identification result by quantifying the degree of matching between the currently detected local feature scale and the overall feature scale and the reference scale. This detection confidence level can intuitively determine whether the current fault pattern recognition is affected by scale deformation caused by changes in the drone's shooting angle and illumination, providing a quantitative basis for the verification and confirmation of subsequent fault detection results, and further improving the accuracy of fault detection.

[0112] To achieve the above objectives, it is first necessary to calculate the similarity between the local feature scale and the reference local feature scale, which is then used as the local similarity. This involves obtaining the current local feature scale extracted in step S200, such as the average scale of multiple insulator discs, and the reference local feature scale determined in step S300. A similarity calculation method suitable for the scale data is then used, such as cosine similarity or Euclidean distance normalization. For example, if cosine similarity is used, the current local feature scale and the reference local feature scale are considered as two vectors, and the similarity is obtained by calculating the cosine of the angle between the vectors. If Euclidean distance normalization is used, the Euclidean distance between the two is first calculated, and then the distance value is normalized to the [0,1] interval. The normalized value is the local similarity; the closer the value is to 1, the higher the local scale matching degree.

[0113] Next, the similarity between the overall feature scale and the reference overall feature scale needs to be calculated as the overall similarity. That is, the current overall feature scale extracted in step S200, such as the overall scale of the insulator, and the reference overall feature scale determined in step S300 are obtained, and calculated using the same method as for calculating local similarity. For example, when using the cosine similarity method, the current overall feature scale and the reference overall feature scale are converted into vectors and the cosine value is calculated; when using the Euclidean distance normalization method, the Euclidean distance between the two is calculated first and then normalized, and the result is the overall similarity. The closer the value is to 1, the higher the overall scale matching degree.

[0114] Finally, the local similarity and overall similarity need to be fused to obtain the detection confidence.

[0115] This involves using a weighted fusion approach to handle local and global similarity. Based on the actual needs of transmission line fault detection, reasonable weights are assigned to local and global similarity (e.g., local detail faults are more dependent on local scale, while overall structural faults are more dependent on overall scale; for example, a weight of 0.6 is set for local scale similarity and 0.4 for global scale similarity, with a sum of 1). The local similarity is multiplied by its corresponding local scale similarity weight, and the global similarity is multiplied by its corresponding global scale similarity weight. The two products are then added together, and the sum is the detection confidence score. For example, if the local similarity is 0.9 and the global similarity is 0.8, calculated using the above weights, the detection confidence score = 0.9 × 0.6 + 0.8 × 0.4 = 0.54 + 0.32 = 0.86. The closer this value is to 1, the higher the reliability of the current fault identification result.

[0116] In step S300 of this application embodiment, self-verification is performed on the local feature scale and the global feature scale to correct the detection confidence and obtain a corrected confidence, including:

[0117] Based on the local feature scale, the estimated overall feature scale is calculated;

[0118] Based on the overall feature scale, the inferred local feature scale is calculated;

[0119] Based on the local feature scale, the global feature scale, the inferred global feature scale, and the inferred local feature scale, self-verification is performed to correct the detection confidence and obtain the corrected confidence.

[0120] In step S300 of this application embodiment, the purpose of the above-mentioned subdivision step is to discover possible deviations in the current scale extraction through logical deduction and mutual verification between local feature scale and global feature scale, and then correct the initial detection confidence to obtain a corrected confidence that is more in line with the actual fault identification situation, so as to avoid inaccurate confidence evaluation due to single scale extraction error and further improve the reliability of fault detection results.

[0121] To achieve the above objectives, it is first necessary to calculate the estimated global feature scale based on the local feature scale, and then calculate the estimated local feature scale based on the global feature scale.

[0122] Specifically, to calculate the overall feature scale, it is necessary to first determine the fixed correlation parameters between the local and overall parts of the target component, such as the number of standard umbrella discs of the insulator, assuming it is 10. Then, obtain the current local feature scale extracted in step S200, such as the average width of a single umbrella disc of the insulator. Multiply the local feature scale by the fixed correlation parameter, such as 10, to obtain the calculated overall feature scale, that is, the calculated overall length of the insulator.

[0123] The local feature scale is calculated based on the fixed correlation parameters mentioned above, such as 10 umbrella discs. The current overall feature scale extracted in step S200 is obtained, such as the actual overall length of the insulator. The overall feature scale is divided by the fixed correlation parameter, such as 10, to obtain the calculated local feature scale, that is, the calculated average width of a single umbrella disc of the insulator.

[0124] Then, based on the local feature scale, the global feature scale, the inferred global feature scale, and the inferred local feature scale, self-verification is required to correct the detection confidence and obtain the corrected confidence.

[0125] In step S300 of this application embodiment, self-verification is performed based on the local feature scale, the global feature scale, the inferred global feature scale, and the inferred local feature scale to correct the detection confidence and obtain the corrected confidence, including:

[0126] The similarity between the local feature scale and the inferred local feature scale is calculated as the local verification score.

[0127] The similarity between the overall feature scale and the inferred overall feature scale is calculated as the overall verification score.

[0128] Based on the local and global verification scores, a verification score is calculated and a correction coefficient is obtained.

[0129] The detection confidence level is corrected by using the correction coefficient to obtain the corrected confidence level.

[0130] Specifically, it is first necessary to calculate the similarity between the local feature scale and the inferred local feature scale as the local verification score, and then calculate the similarity between the overall feature scale and the inferred overall feature scale as the overall verification score.

[0131] Calculating the local verification score requires obtaining the actual extracted local feature scale and the inferred local feature scale. Using a method consistent with the scale similarity calculation, such as cosine similarity, the similarity between the two is calculated. This similarity is the local verification score. The closer the value is to 1, the higher the matching degree between the actual local scale and the inferred local scale, and the more reasonable the local scale extraction.

[0132] To calculate the overall verification score, it is necessary to obtain the actual extracted overall feature scale and the inferred overall feature scale. Using the same similarity calculation method, such as cosine similarity, the similarity between the two is calculated to obtain the overall verification score. The closer the value is to 1, the higher the matching degree between the actual overall scale and the inferred overall feature scale, and the more reasonable the overall scale extraction is.

[0133] Next, the local and overall verification scores need to be fused to obtain the verification score, and then the correction coefficient needs to be calculated. Specifically, the verification score can be calculated using a weighted fusion method. Based on the importance of the local and overall verification scores in the scale rationality assessment, the local verification score is multiplied by its corresponding weight, and the overall verification score is multiplied by its corresponding weight. For example, both weights can be set to 0.5, or adjusted according to the component characteristics. Then, the products of the two are added together to obtain the verification score, which ranges from [0,1]. For example, assuming a local verification score of 0.9, an overall verification score of 0.8, and both weights of 0.5, then the verification score = 0.9 × 0.5 + 0.8 × 0.5 = 0.85.

[0134] Then, calculate the correction coefficient: calculate the correction coefficient according to the formula Correction coefficient = 1 - Verification degree. If the verification degree is 0.85, the correction coefficient = 1 - 0.85 = 0.15; if the verification degree is as low as 0.6, the correction coefficient is 0.4, which reflects that the greater the scale extraction bias, the larger the correction coefficient.

[0135] Finally, the detection confidence level needs to be corrected using the correction coefficient to obtain the corrected confidence level.

[0136] For example, the corrected confidence level can be calculated as: Corrected Confidence Level = Detection Confidence Level × (1 - Correction Coefficient). The formula can also be adjusted according to actual needs. The core logic is that the larger the correction coefficient, the more the initial confidence level is reduced. For instance, if the initial detection confidence level is 0.86 and the correction coefficient is 0.15, then the corrected confidence level = 0.86 × (1 - 0.15) = 0.86 × 0.85 = 0.731; if the correction coefficient is 0.4 and the initial detection confidence level is 0.86, then the corrected confidence level = 0.86 × 0.6 = 0.516, thus achieving a reasonable correction to the initial confidence level.

[0137] In step S300 of this application embodiment, confidence level labeling is also required to complete fault detection, including: using the corrected confidence level to label the identified fault mode to complete fault detection.

[0138] The corrected confidence level obtained in step S300 is associated and bound with the fault mode identified in step S200. While outputting the fault mode, the corresponding corrected confidence level is marked. For example, if the insulator is broken, the corrected confidence level is 0.82. This clarifies the reliability of the fault identification result and completes the entire transmission line fault detection process.

[0139] Example 2, as Figure 2 As shown, based on the same inventive concept as the transmission line fault detection method based on the RKM-YOLO algorithm provided in Embodiment 1, this embodiment of the invention also provides a transmission line fault detection system based on the RKM-YOLO algorithm, including:

[0140] The agent acquisition module 11 is used to acquire a transmission line fault detection agent constructed based on the RKM-YOLO algorithm;

[0141] The fault mode recognition module 12 is used to acquire the transmission line image collected during the current transmission line inspection, input the transmission line image into the transmission line fault detection intelligent agent, output the identified fault mode, and perform local feature scale extraction and global feature scale extraction on the transmission line image.

[0142] The confidence labeling module 13 is used to obtain the baseline local feature scale and baseline global feature scale corresponding to the identified fault mode within a historical time period, analyze the similarity with the baseline local feature scale and baseline global feature scale as the detection confidence, perform self-verification on the local feature scale and global feature scale, correct the detection confidence to obtain the corrected confidence, perform confidence labeling, and complete the fault detection.

[0143] Furthermore, the agent acquisition module 11 includes the following execution steps:

[0144] Based on the YOLOv8 model, the main architecture of the transmission line fault detection agent is constructed.

[0145] Within the main architecture, the RFCAConv module is introduced to first perform group convolution on the input features to extract receptive field features;

[0146] The original C2f module in YOLOv8 is replaced by the K3_RFCA module. The K3_RFCA module adopts a dual parallel 1×1 convolutional branch structure and integrates the RFCAConv module. Shallow and deep features are fused through residual connections.

[0147] The main architecture introduces a multi-scale multi-dimensional collaborative attention module (MMCA), a multi-scale convolution and spatial-channel interaction mechanism, extracts multi-scale features through depthwise separable convolution, and adopts an adaptive weighting strategy to enhance the representation of key regions.

[0148] Based on the aforementioned main architecture, RFCAConv module, and multi-scale multi-dimensional collaborative attention module MMCA, a transmission line fault detection agent is obtained through supervised training.

[0149] Furthermore, the fault mode identification module 12 includes the following execution steps:

[0150] Acquire images of the transmission lines currently being inspected;

[0151] The image of the power transmission line is input into the power transmission line fault detection intelligent agent, and the identified fault mode is output.

[0152] A transmission line scale extractor is obtained, wherein the transmission line scale extractor includes a local scale extraction branch and a global scale extraction branch;

[0153] The image of the transmission line is input into the local scale extraction branch and the global scale extraction branch, and the local feature scale and the global feature scale are output. The local feature scale is the average scale of multiple sheds on the insulator, and the global feature scale is the scale of the insulator.

[0154] The transmission line scale extractor includes:

[0155] Based on historical inspection data of transmission lines, a set of sample transmission line images is collected. The local feature scale and global feature scale of each sample transmission line image are labeled to obtain the sample local feature scale set and the sample global feature scale set.

[0156] Based on convolutional neural networks, local scale extraction branches and global scale extraction branches are constructed.

[0157] Using the sample transmission line image set as input, and the sample local feature scale set and sample global feature scale set as output, the local scale extraction branch and the global scale extraction branch are trained under supervision, and a transmission line scale extractor is obtained after convergence.

[0158] Furthermore, the confidence labeling module 13 includes the following execution steps:

[0159] Acquire all historical images of transmission lines within a historical period that were obtained by the transmission line fault detection agent to identify the identified fault modes and verify their correctness.

[0160] Extract historical local feature scale sets and historical global feature scale sets from all historical transmission line images;

[0161] The occurrence rate of each type of historical local feature scale and historical overall feature scale in the historical local feature scale set and the historical overall feature scale set is statistically analyzed. The historical local feature scale and historical overall feature scale with the highest occurrence rate are selected as the benchmark local feature scale and benchmark overall feature scale.

[0162] Calculate the similarity between the local feature scale and the reference local feature scale, and use it as the local similarity.

[0163] Calculate the similarity between the overall feature scale and the benchmark overall feature scale, and use it as the overall similarity.

[0164] The local similarity and overall similarity are fused to obtain the detection confidence score.

[0165] Based on the local feature scale, the estimated overall feature scale is calculated;

[0166] Based on the overall feature scale, the inferred local feature scale is calculated;

[0167] Based on the local feature scale, the global feature scale, the inferred global feature scale, and the inferred local feature scale, self-verification is performed to correct the detection confidence and obtain the corrected confidence.

[0168] Specifically, based on the local feature scale, the global feature scale, the inferred global feature scale, and the inferred local feature scale, self-verification is performed to correct the detection confidence and obtain the corrected confidence, including:

[0169] The similarity between the local feature scale and the inferred local feature scale is calculated as the local verification score.

[0170] The similarity between the overall feature scale and the inferred overall feature scale is calculated as the overall verification score.

[0171] Based on the local and global verification scores, a verification score is calculated and a correction coefficient is obtained.

[0172] The detection confidence level is corrected by using the correction coefficient to obtain the corrected confidence level.

[0173] This includes confidence level labeling and fault detection, including:

[0174] The identified fault patterns are labeled using the corrected confidence level to complete fault detection.

[0175] It should be noted that the descriptions of each embodiment in the above embodiments have different focuses. For parts that are not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0176] Those skilled in the art will understand that embodiments of the present invention can provide methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0177] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, as well as combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded computer, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0178] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0179] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0180] Although preferred embodiments of the invention have been described, those skilled in the art, once they have learned the basic inventive concept, can make other changes and modifications to these embodiments.

[0181] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of this invention and its equivalents, this invention also intends to include these modifications and variations.

Claims

1. A transmission line fault detection method based on the RKM-YOLO algorithm, characterized in that, The method includes: Obtain a transmission line fault detection intelligent agent constructed based on the RKM-YOLO algorithm; The system acquires images of the transmission lines currently being inspected, inputs these images into the transmission line fault detection agent, outputs the identified fault modes, and performs local and global feature scale extraction on the images. Obtain the baseline local feature scale and baseline global feature scale corresponding to the identified fault mode within a historical time period, analyze the similarity with the baseline local feature scale and baseline global feature scale as the detection confidence, perform self-verification on the local feature scale and global feature scale, correct the detection confidence to obtain the corrected confidence, perform confidence labeling, and complete the fault detection. The process includes self-verification of the local feature scale and the global feature scale, and correcting the detection confidence to obtain a corrected confidence, including: Based on the local feature scale, the estimated global feature scale is calculated; Based on the overall feature scale, the inferred local feature scale is calculated; Based on the local feature scale, the global feature scale, the inferred global feature scale, and the inferred local feature scale, self-verification is performed to correct the detection confidence and obtain the corrected confidence.

2. The transmission line fault detection method based on the RKM-YOLO algorithm according to claim 1, characterized in that, Obtain a transmission line fault detection agent constructed based on the RKM-YOLO algorithm, including: Based on the YOLOv8 model, the main architecture of the transmission line fault detection agent is constructed. Within the main architecture, the RFCAConv module is introduced to first perform group convolution on the input features to extract receptive field features; The original C2f module in YOLOv8 is replaced by the K3_RFCA module. The K3_RFCA module adopts a dual parallel 1×1 convolutional branch structure and integrates the RFCAConv module. Shallow and deep features are fused through residual connections. The main architecture introduces a multi-scale multi-dimensional collaborative attention module (MMCA), a multi-scale convolution and spatial-channel interaction mechanism, extracts multi-scale features through depthwise separable convolution, and adopts an adaptive weighting strategy to enhance the representation of key regions. Based on the aforementioned main architecture, RFCAConv module, and multi-scale multi-dimensional collaborative attention module MMCA, a transmission line fault detection agent is obtained through supervised training.

3. The transmission line fault detection method based on the RKM-YOLO algorithm according to claim 1, characterized in that, The system acquires images of the transmission lines currently being inspected, inputs these images into the transmission line fault detection agent, outputs identified fault patterns, and performs local and global feature scale extraction on the images, including: Acquire images of the transmission lines currently being inspected; The image of the power transmission line is input into the power transmission line fault detection intelligent agent, and the identified fault mode is output. A transmission line scale extractor is obtained, wherein the transmission line scale extractor includes a local scale extraction branch and a global scale extraction branch; The image of the transmission line is input into the local scale extraction branch and the global scale extraction branch, and the local feature scale and the global feature scale are output. The local feature scale is the average scale of multiple sheds on the insulator, and the global feature scale is the scale of the insulator.

4. The transmission line fault detection method based on the RKM-YOLO algorithm according to claim 3, characterized in that, The transmission line scale extractor includes: Based on historical inspection data of transmission lines, a set of sample transmission line images is collected. The local feature scale and global feature scale of each sample transmission line image are labeled to obtain the sample local feature scale set and the sample global feature scale set. Based on convolutional neural networks, local scale extraction branches and global scale extraction branches are constructed. Using the sample transmission line image set as input, and the sample local feature scale set and sample global feature scale set as output, the local scale extraction branch and the global scale extraction branch are trained under supervision, and a transmission line scale extractor is obtained after convergence.

5. The transmission line fault detection method based on the RKM-YOLO algorithm according to claim 1, characterized in that, Obtaining the baseline local feature scale and baseline global feature scale corresponding to the identified fault modes within a historical time period includes: Acquire all historical images of transmission lines within a historical period that were obtained by the transmission line fault detection agent to identify the identified fault modes and verify their correctness. Extract historical local feature scale sets and historical global feature scale sets from all historical transmission line images; The occurrence rate of each type of historical local feature scale and historical overall feature scale in the historical local feature scale set and the historical overall feature scale set is statistically analyzed. The historical local feature scale and historical overall feature scale with the highest occurrence rate are selected as the benchmark local feature scale and benchmark overall feature scale.

6. The transmission line fault detection method based on the RKM-YOLO algorithm according to claim 1, characterized in that, Analyzing the similarity with the benchmark local feature scale and the benchmark global feature scale, as a detection confidence score, includes: Calculate the similarity between the local feature scale and the reference local feature scale, and use it as the local similarity. Calculate the similarity between the overall feature scale and the benchmark overall feature scale, and use it as the overall similarity. The local similarity and overall similarity are fused to obtain the detection confidence score.

7. The transmission line fault detection method based on the RKM-YOLO algorithm according to claim 1, characterized in that, Based on the local feature scale, global feature scale, inferred global feature scale, and inferred local feature scale, self-verification is performed to correct the detection confidence and obtain the corrected confidence, including: The similarity between the local feature scale and the inferred local feature scale is calculated as the local verification score. The similarity between the overall feature scale and the inferred overall feature scale is calculated as the overall verification score. Based on the local and global verification scores, a verification score is calculated and a correction coefficient is obtained. The detection confidence level is corrected by using the correction coefficient to obtain the corrected confidence level.

8. The transmission line fault detection method based on the RKM-YOLO algorithm according to claim 1, characterized in that, Perform confidence level labeling and complete fault detection, including: The identified fault patterns are labeled using the corrected confidence level to complete fault detection.

9. A transmission line fault detection system based on the RKM-YOLO algorithm, characterized in that, The system is used to implement the transmission line fault detection method based on the RKM-YOLO algorithm as described in any one of claims 1-8, including: The agent acquisition module is used to acquire a transmission line fault detection agent constructed based on the RKM-YOLO algorithm; The fault mode recognition module is used to acquire images of the transmission line currently being inspected, input the images of the transmission line into the transmission line fault detection agent, output the identified fault mode, and perform local feature scale extraction and global feature scale extraction on the images of the transmission line. The confidence labeling module is used to obtain the baseline local feature scale and baseline global feature scale corresponding to the identified fault mode within a historical time period, analyze the similarity with the baseline local feature scale and baseline global feature scale as the detection confidence, perform self-verification on the local feature scale and global feature scale, correct the detection confidence to obtain the corrected confidence, perform confidence labeling, and complete the fault detection.

Images