Intelligent real-time detection method and system for hidden danger of distribution line, and medium

Abstract

The invention relates to an intelligent real-time detection method and system for hidden dangers of a distribution line and a medium. The method comprises the following steps: processing multi-modal data collected by an unmanned aerial vehicle; a multi-scale regional attention and dynamic feeling network is designed, a plurality of regional attention modules are combined with local details and global modes of line defect features to carry out feature extraction, information complementation is carried out between high-resolution features and low-resolution features, and classification of defect types is realized; a dynamic receptive field matching module is adopted to optimize a defect bounding box, and a multi-dimensional feature integration module is used to form comprehensive feature representation; a multi-scale regional attention and dynamic feeling network is embedded into unmanned aerial vehicle equipment, and collected data is processed in real time in combination with a calculation unit of the unmanned aerial vehicle, so that a function of detecting hidden dangers while flying is realized. Through the global similarity enhancement denoising technology and the multi-scale feature extraction strategy, the detection precision and adaptability are improved, and the safety and reliability of the distribution line are ensured.

Description

Technical Field

[0001] This application relates to the field of unmanned aerial vehicle (UAV) technology, and more specifically, to a method, system, and medium for intelligent real-time detection of potential hazards in power distribution lines based on UAV data. Background Technology

[0002] Power distribution lines are a crucial component of the power system, and their safe and stable operation directly impacts the system's reliability. Traditional power distribution line inspections rely primarily on manual labor, resulting in low efficiency, high costs, and significant risks. With the rapid development of drone technology, drones are increasingly being used for power distribution line inspections, offering advantages such as high flexibility, low cost, and good safety. However, effectively processing the multimodal data collected by drones and detecting potential defects in power distribution lines in real time remains a challenge. Summary of the Invention

[0003] The purpose of this application is to provide a method, system, and medium for intelligent real-time detection of hidden dangers in power distribution lines based on UAV data. This method can efficiently process multimodal data collected by UAVs and detect hidden defects in power distribution lines in real time, thereby improving inspection efficiency and accuracy.

[0004] To achieve the above objectives, this application provides the following technical solution:

[0005] In a first aspect, embodiments of this application provide an intelligent real-time detection method for potential hazards in power distribution lines, comprising the following steps:

[0006] Step 1: Process the multimodal data collected by the UAV, accurately label the collected image and video data, mark the potential defects in the power distribution lines, and form a high-quality defect sample library; divide the sample library into training set, validation set and test set;

[0007] Step 2: Design a multi-scale region attention and dynamic receptive network. Use multiple region attention modules to extract features by combining local details and global patterns of line defect features. Complementary information is achieved between high-resolution and low-resolution features to classify defect types. Use a dynamic receptive field matching module to optimize the defect bounding box according to the size and shape changes of the defect region. Use a multi-dimensional feature integration module to integrate features from different scales and depths to form a comprehensive feature representation.

[0008] Step 3: Embed the multi-scale regional attention and dynamic perception network into the UAV device, and combine it with the UAV's computing unit to process the collected data in real time, so as to realize the function of detecting hidden dangers while flying.

[0009] In step 1, the image and video data collected by the UAV are subjected to noise reduction preprocessing to improve data quality. Professional annotation tools are used to accurately annotate the preprocessed image and video data to mark the potential defects in the power distribution lines. The annotated data is organized into a defect sample library and the samples are classified according to defect type and severity.

[0010] The denoising preprocessing operation for the image and video data collected by the UAV specifically employs global similarity enhancement denoising.

[0011] The formula for global similarity enhancement and denoising is shown in equation (1):

[0012] (1)

[0013] in, It is a search window, that is, in the location An area around It is a weighting function. Indicates position pixel values ​​relative to position The contribution of the noise reduction value, It is a normalization constant, ensuring that the sum of the weights is 1.

[0014] In step 2, the region attention module receives multi-scale feature maps generated from different feature layers to ensure comprehensive coverage of local details and global patterns. It divides the input feature map into multiple regions of fixed size. For each region, a self-attention mechanism is used to model the feature relationships within the region and calculate the weights between feature channels. Global average pooling is introduced to extract global context features. An attention mechanism is used to fuse features within the region and global features to ensure the consistency of defect features in the overall scene. All weighted region features are restored to the original feature map size to generate the final weighted feature map. The formula for the region attention module is shown in equation (2).

[0015] (2)

[0016] in, This represents the attention weights within the region. Represents global context features, It is a fusion factor that balances local and global features.

[0017] The multi-scale feature map is extracted using a scale-adaptive feature extraction strategy, the formula of which is shown in equation (3):

[0018] (3)

[0019] in, Represents the kernel size. Represents the void ratio, and and It is dynamically selected based on the size and content of the input feature map.

[0020] In step 2, the formulas for the dynamic receptive field matching module are shown in equations (4) and (5):

[0021] (4)

[0022] (5)

[0023] in, It is an adjustable scaling factor. It is the input image In position Local features are typically obtained from the output of convolutional layers. These parameters are obtained through optimization by the dynamic receptive field module and the regression network.

[0024] In step 2, in order to enable feature maps from different scales to be fused in the same spatial dimension, the multi-dimensional feature integration module first needs to use convolution operations to adjust the number of channels and spatial resolution of each feature map. Each feature map at each scale will have its number of channels adjusted by a 1x1 convolution to make the number of channels of each feature map consistent. Since the spatial resolution of the input feature maps is different, these feature maps need to be spatially aligned to make them have the same size. Bilinear interpolation is used to interpolate each feature map, and a weighting coefficient is applied to the feature maps at each scale for weighted fusion to generate the final detection image.

[0025] Secondly, embodiments of this application provide an intelligent real-time detection system for potential hazards in power distribution lines. The system includes: a memory and a processor, a program for the intelligent real-time detection method for potential hazards in power distribution lines, and the steps of the intelligent real-time detection method for potential hazards in power distribution lines implemented when the program for the intelligent real-time detection method for potential hazards in power distribution lines is executed by the processor.

[0026] Thirdly, embodiments of this application provide a computer-readable storage medium storing program code, which, when executed by a processor, implements the steps of the intelligent real-time detection method for hidden dangers in power distribution lines as described above.

[0027] Compared with existing technologies, the advantages of this invention are: by using a multi-scale region attention and dynamic perception network, combined with local details and global patterns for feature extraction, and optimizing defect bounding boxes, high-precision detection is achieved. The system is embedded in UAV equipment, supporting on-flight detection, and combines time-series data processing and global similarity enhancement denoising technology to identify potential hazards and provide feedback in real time. This method improves detection efficiency, accuracy, and adaptability, ensuring the safety and reliability of power distribution lines. Attached Figure Description

[0028] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments of this application will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0029] Figure 1 This is the overall method flowchart of this application;

[0030] Figure 2 This is a diagram illustrating the overall structure of the multi-scale regional attention and dynamic perception network in this application.

[0031] Figure 3 : This is a diagram of the region attention module in this application;

[0032] Figure 4 : This is a diagram of the dynamic receptive field matching module in this application;

[0033] Figure 5 This is a diagram of the multidimensional feature integration module of this application;

[0034] Figure 6 This is a screenshot of the backend user interface of the intelligent real-time detection method for hidden dangers in power distribution lines described in this application. Detailed Implementation

[0035] The technical solutions of the embodiments of this application will now be described with reference to the accompanying drawings. It should be noted that similar reference numerals and letters in the following drawings indicate similar items; therefore, once an item is defined in one drawing, it does not need to be further defined and explained in subsequent drawings.

[0036] The terms “comprising,” “including,” or any other variations thereof are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase “comprising one…” does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0037] The terms “first,” “second,” etc., are used only to distinguish one entity or operation from another, and should not be construed as indicating or implying relative importance, nor as requiring or implying any such actual relationship or order between these entities or operations.

[0038] like Figures 1 to 6 As shown, the present invention provides an intelligent real-time detection method for potential hazards in power distribution lines, comprising the following steps:

[0039] Step 1: Process the multimodal data collected by the drone, use professional annotation tools to accurately annotate the collected image and video data, mark the potential defects in the power distribution lines (such as cracks, damage, foreign object intrusion, etc.), and form a high-quality professional defect sample library.

[0040] In this step, the image and video data collected by the UAV are subjected to denoising preprocessing to improve data quality. The denoising operation includes global similarity enhancement denoising, the core of which is to identify and utilize similar small blocks in the image to reconstruct the value of each pixel. For each pixel in the image, a search window is first defined around it. Within this window, multiple pixel blocks similar to the block where the pixel is located are searched. These similar blocks may be located anywhere in the image, even far from the center pixel. The similarity is usually calculated based on the Euclidean distance between pixel blocks or other similarity measures. For each found similar block, its weight with the center pixel block is calculated. The weight depends on the similarity between them; the higher the similarity, the greater the weight. Commonly used weight functions can be Gaussian functions or other decay functions to ensure that blocks with high similarity contribute more to the final result. The pixel value of each similar block is multiplied by its corresponding weight, and then these weighted pixel values ​​are averaged to estimate the denoised value of the center pixel. This process can be expressed as formula (1):

[0041] (1)

[0042] in, It is a search window, that is, in the location An area surrounding it. It is a weighting function that represents the position. pixel values ​​relative to position The contribution of the noise reduction value. It is a normalization constant, ensuring that the sum of the weights is 1. Weight function Usually based on and The similarity between them can be calculated using formula (2):

[0043] (2)

[0044] in, It is a location and The Euclidean distance between pixel values. These are filtering parameters that control the noise reduction intensity.

[0045] Step 2: Design a multi-scale region attention and dynamic receptive network. Utilize multiple region attention modules to extract features from local details (e.g., crack edges, foreign object outlines) and global patterns (e.g., insulator distribution, conductor orientation) of line defects. Information complementarity is achieved between high-resolution and low-resolution features to classify defect types, including cracks and breaks. A dynamic receptive field matching module is employed to optimize the defect bounding box based on changes in the size and shape of the defect region. A multi-dimensional feature integration module integrates features from different scales and depths to form a comprehensive feature representation.

[0046] In this step, the multi-scale regional attention and dynamic perception network is discussed. Figure 2 It includes a scale-adaptive feature extractor, a region attention module, a dynamic receptive field matching module, and a multi-dimensional feature integration module.

[0047] Scale-adaptive feature extractors are used to extract multi-scale feature maps. Based on the size and complexity of the input features, they dynamically adjust the size and dilation rate of the convolutional kernels to adapt to different receptive fields and capture richer local and global information. When the input image passes through the network, the network selects an appropriate convolutional kernel size and dilation rate based on the scale of the input feature map of the current layer, thereby obtaining receptive fields at different scales. This dynamic selection mechanism can determine the optimal receptive field configuration through parameter learning of the convolutional operation, combined with the size and content of the input features. As shown in Equation 3:

[0048] (3)

[0049] in, Represents the kernel size (optional values ​​are 3, 5, and 7). Represents the void ratio (optional values ​​are 1, 2, and 3), and and It is dynamically selected based on the size and content of the input feature map.

[0050] See the area attention module. Figure 3 The region attention module aims to extract key region information for defect classification from input features. By combining local region features and global contextual information, it enhances feature representation capabilities and adapts to diverse defect features such as cracks and breaks. It receives multi-scale feature maps generated from different feature layers to ensure comprehensive coverage of local details and global patterns. The input feature map is divided into multiple fixed-size regions. For each region, a self-attention mechanism is used to model the feature relationships within the region and calculate the weights between feature channels. An attention mechanism is used to fuse regional and global features to ensure consistency of defect features across the overall scene. Features within each region are redistributed according to the calculated weights. Ideally, a weighted feature map is output, highlighting defect-related regions and suppressing irrelevant regions. As shown in Equation 4:

[0051] (4)

[0052] in, This indicates the final weighted feature map in The value at that location, This represents the attention weights within the region. It is the input feature map at the location eigenvalues ​​at that location Represents global context features, It is a fusion factor that balances local and global features.

[0053] The features calculated using region attention are input into the dynamic receptive field matching module to improve the defect detection system's adaptability to significant changes in the size, shape, and distribution of defect regions, thus optimizing the localization and shape of the defect bounding box. See the dynamic receptive field matching module for details. Figure 4 The dynamic receptive field matching module is an adaptive mechanism designed to dynamically adjust the size and shape of the receptive field based on local information from the input feature map. In this way, the network can more accurately capture details of defect regions, especially when the shape and size of the defect change, enabling it to adjust the receptive field promptly to adapt to these changes. The size of the input feature map is... At a certain location in the feature map The corresponding receptive field size is It can be dynamically changed through the receptive field adjustment function, as shown in Equation 5:

[0054] (5)

[0055] in, It is an adjustable scaling factor. It is the input image In position The local features are obtained from the output of the convolutional layer. Bounding box optimization mainly involves adjusting the parameters of the detection box (such as position, width, and height) to more accurately enclose the defect region. The introduction of the dynamic receptive field matching module allows the bounding box optimization to be adjusted in real time after sensing changes in the shape and size of the defect. Assuming the original bounding box... ,in and The center coordinates of the bounding box and The bounding box is defined by its width and height. The process of optimizing the bounding box using the dynamic receptive field module can be performed through the following steps. First, position optimization is performed: the center coordinates of the bounding box are adjusted based on the response of the dynamic receptive field to local features, as shown in Equation 6.

[0056] (6)

[0057] in, and This is the coordinate offset adjusted through dynamic receptive field. Secondly, size optimization is performed: the width and height of the bounding box are adjusted according to the shape and size of the defect, as shown in Equation 7.

[0058] (7)

[0059] in, and These are the width and height of the bounding box, adaptively adjusted based on the receptive field information. Finally, after adaptive adjustment of the defect region using a dynamic receptive field matching module, the final optimized bounding box is obtained. As shown in Equation 8:

[0060] (8)

[0061] in, These parameters are obtained through optimization by the dynamic receptive field module and the regression network.

[0062] See the multi-dimensional feature integration module. Figure 5 To enable the fusion of feature maps from different scales in the same spatial dimension, convolution operations are first used to adjust the number of channels and spatial resolution of each feature map. Each feature map at each scale is adjusted for its number of channels using a 1x1 convolution to ensure consistency. Simultaneously, bilinear interpolation is used to adjust the feature maps to the same spatial resolution. The multi-scale features are then weighted and fused through concatenation, as shown in Equation 9.

[0063] (9)

[0064] in, These are parameters obtained through training, representing the importance of features at different scales.

[0065] Step 3: Design a real-time hazard detection system for UAVs. Embed a multi-scale area attention and dynamic sensing network into the UAV device. Combined with the UAV's computing unit, the system processes the collected data in real time, enabling hazard detection during flight. The dynamic sensing network, combined with time-series data processing, accurately identifies potential hazards or faults during flight through continuous frame analysis. The computing unit includes a high-performance processor, supporting rapid data processing and real-time feedback. Furthermore, by integrating the embedded system with the flight control system, it achieves coordinated optimization of hazard detection and flight control, ensuring flight safety and reliability.

[0066] In this step, the multi-scale region attention and dynamic perception network is trained end-to-end using the refined sample library created in step 1 to optimize network parameters. Cross-entropy loss and region focusing loss are jointly optimized to ensure the model can accurately identify defect regions. The continuous frame analysis refers to the UAV continuously acquiring images of power distribution lines at a fixed frame rate during flight. Real-time images are input into the optimized network for defect identification, generating real-time defect feedback. Since continuous frame analysis generates a large number of images to be processed, the system is equipped with a high-performance computing unit to support rapid data processing and real-time feedback. A high-performance embedded processor (such as NVIDIA Jetson Xavier or Intel Core i7) is used to support parallel computing and deep learning inference. Large-capacity memory (such as 16GB RAM) and high-speed storage (such as SSD) ensure efficient data processing. GPU acceleration technology is used to achieve parallel computing for image processing and deep learning inference. Through optimized algorithms and hardware resource scheduling, the system can complete data processing and feedback within milliseconds. Simultaneously, the embedded system works in conjunction with the flight control system to dynamically adjust flight parameters based on the hazard detection results, ensuring flight safety and reliability. The embedded system dynamically adjusts the drone's flight parameters (such as altitude, speed, and heading) based on the detection results. When a serious hazard is detected, the system triggers emergency actions (such as hovering, returning to home, or landing). The collaborative work between the embedded system and the flight control system further enhances flight safety and reliability, providing strong technical support for the application of drones in power line inspection.

[0067] This application provides an intelligent real-time detection system for potential hazards in power distribution lines. Figure 6The system includes: a memory and a processor, a program for the intelligent real-time detection method for hidden dangers in power distribution lines, and the steps of the intelligent real-time detection method for hidden dangers in power distribution lines implemented by the processor when the program for the intelligent real-time detection method for hidden dangers in power distribution lines is executed.

[0068] This application provides a computer-readable storage medium storing program code. When the program code is executed by a processor, it implements the steps of the intelligent real-time detection method for hidden dangers in power distribution lines as described above.

[0069] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application 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.

[0070] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and 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 processor, 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.

[0071] 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.

[0072] 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 1The steps of the function specified in one or more boxes.

[0073] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0074] Memory may include non-persistent memory in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0075] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0076] The above description is merely an embodiment of this application and is not intended to limit the scope of protection of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A method for intelligent real-time detection of hidden dangers in power distribution lines, characterized in that, Includes the following steps: Step 1: Process the multimodal data collected by the drone, accurately label the collected image and video data, mark the potential defects in the power distribution lines, and form a high-quality defect sample library; The sample database is divided into a training set, a validation set, and a test set; Step 2: Design a multi-scale region attention and dynamic receptive network. Use multiple region attention modules to extract features by combining local details and global patterns of line defect features. Complementary information is achieved between high-resolution and low-resolution features to classify defect types. Use a dynamic receptive field matching module to optimize the defect bounding box according to the size and shape changes of the defect region. Use a multi-dimensional feature integration module to integrate features from different scales and depths to form a comprehensive feature representation. Step 3: Embed the multi-scale regional attention and dynamic perception network into the UAV device, and combine it with the UAV's computing unit to process the collected data in real time, so as to realize the function of detecting hidden dangers while flying.

2. The intelligent real-time detection method for hidden dangers in power distribution lines according to claim 1, characterized in that, In step 1, the image and video data collected by the UAV are subjected to noise reduction preprocessing to improve data quality. Professional annotation tools are used to accurately annotate the preprocessed image and video data to mark the potential defects in the power distribution lines. The annotated data is organized into a defect sample library and the samples are classified according to defect type and severity.

3. The intelligent real-time detection method for hidden dangers in power distribution lines according to claim 2, characterized in that, The denoising preprocessing operation for the image and video data collected by the UAV specifically employs global similarity enhancement denoising. The formula for global similarity enhancement and denoising is shown in equation (1): （1） in, It is a search window, that is, in the location An area around It is a weighting function. Indicates position pixel values ​​relative to position The contribution of the noise reduction value, It is a normalization constant, ensuring that the sum of the weights is 1.

4. The intelligent real-time detection method for hidden dangers in power distribution lines according to claim 1, characterized in that, In step 2, the region attention module receives multi-scale feature maps generated from different feature layers, ensuring comprehensive coverage of local details and global patterns. It divides the input feature map into multiple fixed-size regions. For each region, a self-attention mechanism is used to model the feature relationships within the region, calculate the weights between feature channels, extract global contextual features by introducing global average pooling, and fuse regional and global features using an attention mechanism. To ensure the consistency of defect features across the overall scene, all weighted region features are restored to the original feature map size to generate the final weighted feature map. The formula for the region attention module is shown in equation (2). （2） in, This indicates the final weighted feature map in The value at that location, This represents the attention weights within the region. It is the input feature map at the location eigenvalues ​​at that location Represents global context features, It is a fusion factor that balances local and global features.

5. The intelligent real-time detection method for hidden dangers in power distribution lines according to claim 4, characterized in that, The multi-scale feature map is extracted using a scale-adaptive feature extraction strategy, the formula of which is shown in equation (3): （3） in, This is the result after performing a convolution operation on the input feature map, where k represents the k-th feature map. Represents the kernel size. Represents the void ratio, and and It is dynamically selected based on the size and content of the input feature map. This indicates the input feature map Perform three different convolution operations and sum their results.

6. The intelligent real-time detection method for hidden dangers in power distribution lines according to claim 1, characterized in that, In step 2, the formulas for the dynamic receptive field matching module are shown in equations (4) and (5): （4） （5） in, It is an adjustable scaling factor. It is the input image In position Local features are typically obtained from the output of convolutional layers. These parameters are obtained through optimization by the dynamic receptive field module and the regression network.

7. The intelligent real-time detection method for hidden dangers in power distribution lines according to claim 1, characterized in that, In step 2, in order to enable feature maps from different scales to be fused in the same spatial dimension, the multi-dimensional feature integration module first needs to use convolution operations to adjust the number of channels and spatial resolution of each feature map. Each feature map at each scale will have its number of channels adjusted by a 1x1 convolution to make the number of channels of each feature map consistent. Since the spatial resolution of the input feature maps is different, these feature maps need to be spatially aligned to make them have the same size. Bilinear interpolation is used to interpolate each feature map, and a weighting coefficient is applied to the feature maps at each scale for weighted fusion to generate the final detection image.

8. A real-time intelligent detection system for potential hazards in power distribution lines, characterized in that, The system includes: a memory and a processor; a program for the intelligent real-time detection method for hidden dangers in power distribution lines; and, when the program for the intelligent real-time detection method for hidden dangers in power distribution lines is executed by the processor, the following steps are performed: Step 1: Process the multimodal data collected by the UAV, accurately label the collected image and video data, mark the potential defects in the power distribution lines, and form a high-quality defect sample library; divide the sample library into training set, validation set and test set; Step 2: Design a multi-scale region attention and dynamic receptive network. Use multiple region attention modules to extract features by combining local details and global patterns of line defect features. Complementary information is achieved between high-resolution and low-resolution features to classify defect types. Use a dynamic receptive field matching module to optimize the defect bounding box according to the size and shape changes of the defect region. Use a multi-dimensional feature integration module to integrate features from different scales and depths to form a comprehensive feature representation. Step 3: Embed the multi-scale regional attention and dynamic perception network into the UAV device, and combine it with the UAV's computing unit to process the collected data in real time, so as to realize the function of detecting hidden dangers while flying.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores program code, which, when executed by a processor, implements the steps of the intelligent real-time detection method for hidden dangers in power distribution lines as described in any one of claims 1 to 7.

Images