Lightweight intelligent inspection method for high-voltage transmission line based on unmanned aerial vehicle terminal

Abstract

The invention relates to a high-voltage transmission line lightweight intelligent inspection method based on an unmanned aerial vehicle terminal, and belongs to the field of intelligent inspection of power equipment, and the method comprises the following steps: 1, collecting an original sample set; 2, performing defect labeling, preprocessing and sample expansion on the original sample set; dividing a training set and a test set, and performing data enhancement on the training set; step 3, constructing an insulator defect and foreign matter detection model, including introducing a cross-space learning multi-scale attention mechanism EMA and a feature channel shuffling enhancement model on the basis of the YOLOv8m network; designing a dual-path guide feature complementary fusion module DG-CFM; a cascade multi-core module CMKM is constructed; 4, performing training and parameter tuning on the model by using the training set; and step 5, deploying the trained detection model to an inspection unmanned aerial vehicle platform with edge computing power, performing testing by using a test set, and performing practical application.

Description

Technical Field

[0001] This invention belongs to the field of intelligent inspection technology for power equipment, and relates to a lightweight intelligent inspection method for high-voltage transmission lines based on a drone terminal. Background Technology

[0002] With the continuous development of modern electricity demand and power systems, the safety and stability of transmission lines have attracted widespread attention. Due to long-term exposure to complex outdoor environments and various factors such as long-term operation, insulators, as a crucial electrical component in power transmission and distribution, are susceptible to corrosion, material aging, and other factors, potentially leading to defects such as spontaneous explosion or breakage. Furthermore, foreign objects such as bird nests and honeycomb deposits may adhere to transmission lines and towers, causing short-circuit faults. Therefore, efficient and accurate defect detection of insulators and attached foreign objects is of great practical significance in ensuring the stability and service life of transmission lines.

[0003] Traditional methods for detecting insulator defects and foreign objects rely heavily on manual inspections, which suffer from low safety and efficiency. Furthermore, the limitations of manual on-site inspections in the field are significant. In recent years, deep learning technology has made significant progress in image processing and object detection, providing new approaches to transmission line defect detection. By utilizing high-definition imaging equipment such as wide-angle and zoom cameras mounted on UAV platforms to efficiently acquire image data of transmission lines and their various components, and training a deep learning object detection model, automatic identification and classification of insulator defects and foreign objects in the acquired images can be achieved, greatly improving the accuracy and efficiency of detection.

[0004] Currently, deep learning object detection algorithms based on convolutional neural networks can be divided into two categories: two-stage detection algorithms and one-stage detection algorithms. Two-stage detection algorithms, mainly including R-CNN, Fast R-CNN, Faster R-CNN, and FPN, have high detection accuracy but are slow, making them difficult to meet the real-time, fast inspection requirements. In contrast, one-stage detection algorithms such as You Only Look Once (YOLO) and Single Shot MultiBox Detector (SSD) can simultaneously complete object recognition and localization, offering faster detection speeds and achieving real-time detection while maintaining high accuracy. YOLOv8 is the object detection algorithm in the YOLO series, featuring numerous technical improvements and optimizations, and is widely used and technically mature in industrial scenarios. This algorithm achieves further lightweighting by introducing a C2f module, abandoning the previous Intersection of Union (IoU) matching or unilateral proportional allocation methods, and using Task-AlignedAssigner matching, significantly improving detection accuracy and robustness. YOLOv8 significantly improves detection accuracy while maintaining speed by integrating deeper and more complex network structures and improved training techniques. During training, it employs various data augmentation techniques and strategies, such as random cropping, color gradation, and mosaic enhancement. These methods effectively expand the dataset and increase the diversity of data samples. This diversity is crucial for improving the model's generalization ability, allowing it to learn broader and more complex scene features, thus enabling accurate and reliable predictions even when encountering new and unseen data.

[0005] Currently, there are two major challenges in using UAV aerial images for insulator defect and foreign object detection in power transmission lines: (1) UAV aerial images of power transmission lines often face challenges such as complex backgrounds, large interference with targets, large differences in defect target scale, and target overlap, resulting in poor generalization ability of detection models, low detection accuracy, and difficulty in accurately detecting small-scale defects; (2) Due to the limited computing and storage capabilities of UAV platforms, most existing deep defect detection models are deployed on cloud servers. After the UAV terminal collects the image, it needs to go through multiple data transmission stages before defect detection can be performed, which increases the computing overhead and results in poor real-time performance, making it unsuitable for real-time monitoring and seriously affecting the normal operation of the power system. Therefore, this method constructs a lightweight, high-precision lightweight multi-defect detection model for high-voltage power transmission lines based on the improved YOLOv8m, which can be directly deployed on UAV platforms. It can quickly detect and monitor insulator defects and attached foreign objects in real time, which is of great significance for ensuring the safe and stable operation of the power system. Summary of the Invention

[0006] In view of this, the purpose of the present invention is to provide a lightweight intelligent inspection method for high-voltage transmission lines based on a drone terminal.

[0007] To achieve the above objectives, the present invention provides the following technical solution: A lightweight intelligent inspection method for high-voltage transmission lines based on a drone terminal includes the following steps: Step 1: Collect images of insulators, bird nests, and honeycomb-like foreign objects during the transmission line inspection process to form the original sample set; Step 2: Label the defects in the original sample set, preprocess the original sample set, and expand the samples using public datasets to form a defect dataset; divide the defect data into training set and test set, perform data augmentation on the training set, and expand the training sample size. Step 3: Construct the insulator defect and foreign object detection model DD-YOLOv8m, including: on the basis of the YOLOv8m network, using the multi-scale attention mechanism EMA and the feature channel shuffling enhancement model to integrate and preserve the target space details; designing the dual-path guided feature complementary fusion module DG-CFM to promote the alignment of shallow positional information and deep semantic information; and constructing the cascaded multi-core module CMKM to alleviate the problem of loss of deep small target features in the network, enhance multi-scale perception and reduce model complexity. Step 4: Train and optimize the parameters of the insulator defect and foreign object detection model DD-YOLOv8m using the training set; Step 5: Deploy the trained detection model to an inspection drone platform with edge computing power, test the model using a test set, and apply it in practice.

[0008] Furthermore, in step 1, when the drone inspects the transmission lines and towers, it takes pictures of insulators, bird nests, and honeycomb-like foreign objects using its onboard camera. Unusable images are removed to obtain a preliminary raw sample set.

[0009] Furthermore, in step 2, various defect features are labeled offline on the original sample set, including insulator spontaneous explosion, bird nest foreign objects, and honeycomb foreign objects. A defect dataset is obtained by fusing the public dataset with the labeled original sample set. The defect dataset includes five types of defects: insulator damage, insulator flashover, insulator spontaneous explosion, bird nest foreign objects, and honeycomb foreign objects. The newly constructed defect dataset is divided into a training set and a test set. Data augmentation operations are performed on the training set by introducing rotation, translation, random cropping, brightness adjustment, and noise addition.

[0010] Furthermore, step 3, based on the YOLOv8m network, utilizes the multi-scale attention mechanism EMA (Effective Multi-Scale Attention) and the feature channel shuffling enhancement model to integrate and preserve the detailed features of the target space. Specifically, this includes replacing the shallow C2f modules in layers 3 and 5 of the backbone network with EC-C2f modules to construct a deep feature reuse chain; the EC-C2f module consists of the multi-scale attention mechanism EMA, feature channel grouping processing, and feature shuffling processing. The Multi-Scale Attention (EMA) mechanism models long-range dependencies, embeds precise location information, captures cross-scale spatial dependencies through parallel sub-networks, and adaptively fuses feature responses from different receptive fields. The feature channel grouping process divides the input feature channels into several subgroups, and extracts one or more groups of features each time. The feature shuffling process is used to exchange information between grouped features. The number of repetitions is consistent with the number of bottleneck layers, thereby promoting feature reuse and reducing the number of parameters and computational complexity.

[0011] Furthermore, the multi-scale attention mechanism (EMA) specifically includes: (1) Feature grouping: Assume the input feature map is The average grouping is performed along the channel dimension, defined as follows:

[0012] (2) Parallel sub-network processing: Attention weights of grouped feature maps are extracted simultaneously for each segment through two parallel sub-networks; in the 1×1 convolution branch, average pooling is performed on the input along the X-axis and average pooling is performed along the Y-axis, and the channels are encoded along the two spatial directions respectively; the 3×3 convolution branch uses a single 3×3 kernel stacked to capture multi-scale feature representations; In the 1×1 convolution branch, the input feature map is first subjected to global average pooling in the horizontal and vertical directions to generate two feature maps with dimensions of C / / G×H×1 and C / / G×1×W, respectively. This obtains long-range dependencies in the horizontal direction and long-range interactions in the spatial direction, while preserving precise positional information in the vertical and horizontal dimensions.

[0013]

[0014] in, For the input feature map, These represent the height and width of the input feature map, respectively. The first c Each channel is at a height H The output at the width W The output at that location, For feature map in widthW Location and height H Pixel value at; The generated feature map of size C / / G×H×1 is transformed, then concatenated with the feature map of size C / / G×1×W. A shared 1×1 convolution is used for further transformation, followed by channel splitting, resulting in two independent tensors. and :

[0015]

[0016] in, These represent two 1×1 convolution transformation functions. This represents the sigmoid activation function; Spatial attention is obtained by multiplying the original grouped input with the obtained attention score; the 3×3 convolution branch captures local cross-channel interactions through 3×3 convolution, expanding the feature space; (3) Cross-spatial learning: The two output tensors of the processed 1×1 convolutional branch and 3×3 convolutional branch are used to encode the global spatial information of the output in the 1×1 convolutional branch using 2D global average pooling and to model the long-range dependency. Before the joint activation mechanism of the channel features, the output of the minimum branch is directly converted into the corresponding dimensional shape, i.e. The expression for 2D global average pooling is as follows:

[0017] The outputs of parallel processing are multiplied by matrix dot products to generate the first spatial attention map. 2D global average pooling is used to encode the global information in the 3×3 convolutional branches. Before the joint activation mechanism for channel features, the global spatial information of the 1×1 branches is converted into its corresponding dimensional shape. Derive the second spatial attention map; Finally, the output feature maps within each group are aggregated using two generated spatial attention weight values ​​to capture pixel-level pairings and highlight the global context of all pixels. After embedding the multi-scale attention mechanism (EMA) into the second convolution of the bottleneck layer of the C2f module in layers 3 and 5 of the backbone network, the original features are preserved through identity mapping connections. The principle is as follows:

[0018] in, This represents the output of the bottleneck layer structure after introducing the coordinate attention mechanism. Indicates the input feature map, This represents an integrated module consisting of two 3×3 convolutions and an EMA attention mechanism.

[0019] Furthermore, the feature channel grouping processing divides the input feature channels into two groups for processing, and the specific principle is as follows:

[0020]

[0021]

[0022]

[0023] in, Indicates the number of channels is n Feature map, Representation of feature map The first in n One channel, This represents the bottleneck layer output after feature channel grouping processing. This represents a module consisting of two 3×3 convolutions, a multi-scale attention mechanism (EMA), and residual connections. This indicates that the feature maps are spliced ​​and aggregated along the channel dimension.

[0024] Furthermore, step 3, which involves designing a dual-path guided feature complementary fusion module (DG-CFM), promotes the alignment of shallow location information with deep semantic information. Specifically, this includes replacing the C2f modules in layers 7 and 9 of the backbone network with the dual-path guided feature complementary fusion module. This module consists of both channel feature guidance and spatial feature guidance, which recalibrate the input features. The specific construction is as follows: (1) Channel splitting: Assume the input feature map is It is split into two parts along the channel dimension, one part contains The channel, another part contains The channel is described as follows:

[0025] in, ; (2) Dual-path guidance: The input consists of a position-preserving branch composed of pointwise convolutions, which preserves the fine-grained spatial structure and positional information from shallow layers in the input features. The output of this branch is represented as... ;Will The input consists of a semantic enhancement branch composed of standard 3×3 convolutions, responsible for extracting contextual semantic features along the channel dimension. The output of this branch is represented as... The transformation process is represented as follows:

[0026] in, A function representing the mapping relationship between the learning space and semantic information. The original spatial location information was preserved. It contains semantic information of the multichannel; (3) Complementary enhancement: Through spatial feature guidance modules, for Weights are assigned to important information at each location, and these weights are mapped to... In this process, its structural awareness is enhanced; through the channel feature guidance module, weights are assigned to important information in each channel, and these weights are mapped to... In this context, enhance its semantic hierarchy; (4) Feature fusion: By multiplying point by point, the feature is fused. Its mapping to Above; by multiplying point by point... Its mapping to Then, the two branches are merged together to obtain the output features. It contains the feature of a dual complementary mapping of spatial and semantic relationships. The calculation is as follows:

[0027] in, This indicates element-wise multiplication. This indicates element-wise addition.

[0028] Furthermore, the spatial feature guidance module first uses a 1×1 convolutional layer to... The number of channels is reduced to one dimension, and then batch normalization is used to obtain spatial information at each location. Finally, a sigmoid activation function layer is used to obtain the key spatial information weights; unique weights are generated in space. The process is described as follows:

[0029] in, express The feature map after dimensionality reduction processing by a 1×1 convolutional layer.

[0030] Furthermore, the channel feature guidance module first uses deep convolutional layers to... Each channel undergoes convolution, followed by global average pooling to obtain global information for each channel. Finally, key information weights are obtained through a sigmoid activation function layer, generating unique weights for each channel. The process is described as follows:

[0031] in, express The feature map obtained after processing by deep convolutional layers.

[0032] Furthermore, step 3, which involves constructing a cascaded multi-kernel module (CMKM), alleviates the problem of feature loss in deep small targets within the network, enhances multi-scale perception, and reduces model complexity. Specifically, this includes replacing the last downsampling convolution in the backbone network with a cascaded multi-kernel module. This cascaded multi-kernel module consists of depthwise convolutional kernels of different sizes, with spatial point convolutions added between these kernels to focus on details at different scales, highlighting multi-scale feature representation. The overall construction of the cascaded multi-kernel module is as follows:

[0033] in, These represent the input and output feature maps, respectively. Indicates the kernel size as Depth convolution; This represents pointwise convolution.

[0034] The beneficial effects of this invention are as follows: The lightweight intelligent inspection method for high-voltage transmission lines based on UAV terminals can efficiently, accurately, and in real-time detect insulator defects, bird nests, and honeycomb foreign objects in complex backgrounds, overlapping targets, and severe obstructions, thus avoiding serious impacts from defects; the designed EC-C2f module ensures that defect location features are continuously preserved and refined during forward propagation by constructing an effective deep feature reuse chain, enhancing the model's ability to integrate and preserve detailed features of the target space, reducing the number of parameters and computational load to optimize the model; the designed dual-path guided feature complementary fusion module promotes effective alignment of shallow location information and deep semantic information, enhancing the ability to locate small target defects; the constructed cascaded multi-core module alleviates the problem of loss of deep small target features in the network, enhances multi-scale perception, reduces model complexity, and improves the accuracy of various defect detections; deploying the model on a low-altitude inspection UAV platform with lightweight computing power enables efficient and real-time detection of insulator defects and foreign objects on high-voltage transmission lines. This invention provides an efficient, intelligent, and real-time monitoring solution for detecting various defects such as insulator defects and foreign matter attachments in high-voltage transmission lines in the power inspection industry. It is of great significance for ensuring the safe and stable operation of the power system and promoting intelligent and automated power inspection.

[0035] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination, or may be learned from practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description

[0036] To make the objectives, technical solutions, and advantages of the present invention clearer, the preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings, wherein: Figure 1 Flowchart of the lightweight intelligent inspection method for high-voltage transmission lines based on a low-altitude unmanned aerial vehicle terminal provided by the present invention; Figure 2 The structure diagram of the DD-YOLOv8m network model for detecting various defects such as insulator defects and foreign objects in high-voltage transmission lines, based on the improved YOLOv8m, provided by this invention; Figure 3 A structural diagram of the efficient multi-scale attention mechanism for cross-space learning provided by this invention; Figure 4 This is a structural diagram of the EC-C2f module provided by the present invention; Figure 5 This is a structural diagram of the dual-path guidance feature complementary fusion module described in this invention; Figure 6 This is a structural diagram of the cascaded multi-core module described in this invention; Figure 7 The image shows the detection results of the DD-YOLOv8m detection model provided by this invention deployed on a low-altitude inspection UAV platform. Detailed Implementation

[0037] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0038] It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Therefore, the drawings only show the components related to the present invention and are not drawn according to the actual number, shape and size of the components in the actual implementation. In the actual implementation, the form, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0039] In the following description, numerous details are explored to provide a more thorough explanation of embodiments of the invention. However, it will be apparent to those skilled in the art that embodiments of the invention may be practiced without these specific details. In other embodiments, well-known structures and devices are shown in block diagram form rather than in detail to avoid obscuring embodiments of the invention.

[0040] Example 1: This invention provides a lightweight intelligent inspection method for high-voltage transmission lines based on a low-altitude unmanned aerial vehicle (UAV) terminal, comprising the following steps: Step 1: Collect images of insulators, bird nests, and honeycomb-like foreign objects during the inspection of transmission lines to form the original sample set.

[0041] Specifically, when drones inspect power transmission lines and towers, they use their high-definition cameras to capture images containing insulators, bird nests, and honeycomb-like foreign objects. Images that are too poor in quality to be used due to shooting errors such as blurriness, overexposure or underexposure, or poor shooting angles are removed, resulting in a preliminary raw sample set.

[0042] Step 2: Use annotation software to properly annotate the defects in the actual collected sample set, use image processing methods to preprocess the images, and use public datasets to expand the samples to form a defect dataset; divide the defect data into training set and test set, and use data augmentation techniques to augment the training set to expand the training sample size.

[0043] Specifically, various defect features, including insulator spontaneous explosion, bird nest foreign objects, and honeycomb foreign objects, are labeled offline on the original sample set. The original sample set is then fused with a public dataset to obtain a richer sample set, including five typical defects: insulator damage, insulator flashover, insulator spontaneous explosion, bird nest foreign objects, and honeycomb foreign objects. The newly constructed sample set is divided into a training set and a test set. Data augmentation operations such as rotation, translation, random cropping, brightness adjustment, and noise addition are introduced into the training set to enrich the morphological and scale representation of the images and improve the reproduction frequency of various defects.

[0044] Step 3: Based on the YOLOv8m model, an improved insulator defect and foreign object detection model, DD-YOLOv8m, is constructed. This includes introducing an efficient multi-scale attention mechanism (EMA) for cross-spatial learning and a feature channel shuffling enhancement model to integrate and preserve the detailed features of the target space; designing a dual-path guided feature complementary fusion module (DG-CFM) to promote the effective alignment of shallow positional information and deep semantic information, thereby enhancing the localization capability of small targets; and constructing a cascaded multi-core module (CMKM) to alleviate the problem of loss of deep small target features in the network, enhance multi-scale perception, and reduce model complexity.

[0045] Specifically, step 3 includes designing a new Cross Stage Partial with 2 Convolutions-fusion (C2f) structure, EC-C2f, which replaces the shallow C2f modules in layers 3 and 5 of the backbone network with EC-C2f modules, constructing a deep feature reuse chain to ensure that defect localization features are continuously preserved and refined during forward propagation, enhancing the model's ability to integrate and preserve detailed features in the target space, and reducing the number of parameters and computational cost to optimize the model; designing a dual-path guided feature complementary fusion module to replace the C2f modules in layers 7 and 9 to promote effective alignment between shallow positional information and deep semantic information, enhancing the ability to localize small targets; and constructing a cascaded multi-core module to replace the last downsampling convolution in the backbone to alleviate the problem of loss of deep small target features in the network, enhance multi-scale perception, reduce model complexity, and improve the accuracy of various defect detections.

[0046] Furthermore, an EMA attention mechanism is embedded in the bottleneck layer of the shallow C2f architecture of the backbone network, enabling the model to focus on defect features and suppress redundant background information, thereby improving detection accuracy. The EMA attention mechanism not only models long-range dependencies but also embeds precise location information, thus strengthening the feature representation of key defect regions. By capturing cross-scale spatial dependencies through parallel sub-networks, it can adaptively fuse feature responses from different receptive fields, improving the perception of defects such as insulator cracks, spontaneous explosions, and foreign matter attachments, suppressing background interference, and enhancing detection robustness in scenarios with viewing angle deviations and partial occlusion. The overall construction of the EMA attention mechanism is as follows: (1) Feature grouping: Assume the input feature map is The average grouping is performed along the channel dimension, defined as follows:

[0047] (2) Parallel sub-network processing: Attention weights of grouped feature maps are extracted simultaneously for each segment through two parallel sub-networks. In the 1×1 convolution branch, average pooling is performed on the input along the X-axis and average pooling is performed along the Y-axis to encode the channels along the two spatial directions respectively; the 3×3 convolution branch uses a single 3×3 kernel stacked to capture multi-scale feature representations.

[0048] In the 1×1 convolution branch, the input feature map is first subjected to global average pooling in the horizontal and vertical directions to generate two feature maps with dimensions of C / / G×H×1 and C / / G×1×W, respectively. This obtains long-range dependencies in the horizontal direction and long-range interactions in the spatial direction, while preserving precise positional information in the vertical and horizontal dimensions.

[0049]

[0050] in, For the input feature map, These represent the height and width of the input feature map, respectively. The first c Each channel is at a height H The output at the width W The output at that location, For feature map in width W Location and height H The pixel value at that location.

[0051] The generated feature map of size C / / G×H×1 is transformed, then concatenated with the feature map of size C / / G×1×W. A shared 1×1 convolution is used for further transformation, followed by channel splitting, resulting in two independent tensors. and :

[0052]

[0053] in, These represent two 1×1 convolution transformation functions. This represents the sigmoid activation function.

[0054] Spatial attention is obtained by multiplying the original grouped input by the obtained attention score. On the other hand, the 3×3 convolution branch captures local cross-channel interactions through 3×3 convolution, expanding the feature space. In this way, EMA not only encodes inter-channel information to adjust the importance of different channels, but also preserves accurate spatial structure information in the channels.

[0055] (3) Cross-spatial learning: A cross-spatial information aggregation method is designed to aggregate richer features across different spatial dimensions. Specifically, after the above processing, two output tensors, a 1×1 convolutional branch and a 3×3 convolutional branch, are obtained. 2D global average pooling is used to encode the global spatial information of the output in the 1×1 convolutional branch and model the long-range dependency. Before the joint activation mechanism of the channel features, the output of the minimum branch is directly converted into the corresponding dimensional shape, i.e. The expression for 2D global average pooling is as follows:

[0056] The outputs of the parallel processing described above are multiplied using matrix dot products to generate the first spatial attention map, which collects spatial information at different scales. Subsequently, 2D global average pooling is used to encode the global information in the 3×3 convolutional branch. Before the joint activation mechanism for channel features, the global spatial information of the 1×1 branch is directly converted into its corresponding dimensional shape. This process derives a second spatial attention map that preserves the entire precise spatial location information. Finally, the output feature maps within each group are aggregated using the two generated spatial attention weights to capture pixel-level pairings and highlight the global context of all pixels.

[0057] When attention mechanisms use learned attention weights to weight features, they can easily filter out some useful but small-scale features. Therefore, this paper embeds the EMA attention mechanism after the second convolution of the bottleneck layer of the C2f module in layers 3 and 5 of the backbone network, and retains the original features through identity mapping connections, which promotes feature reuse when bottleneck layers are stacked. The principle is as follows:

[0058] in, This represents the output of the bottleneck layer structure after introducing the coordinate attention mechanism. Indicates the input feature map, This represents an integrated module consisting of two 3×3 convolutions and an EMA attention mechanism.

[0059] Furthermore, channel shuffling is introduced to improve the original bottleneck layer in the backbone network, designing a new feature reuse bottleneck layer: composed of three parts: EMA attention mechanism, channel grouping processing, and shuffling processing. Although the identity mapping connection method achieves feature reuse, embedding EMA into the bottleneck layer leads to module stacking, resulting in information redundancy and an increase in model parameters, affecting detection performance. An improved feature reuse strategy is proposed: combining EMA, feature channel grouping, and shuffling processing to construct a new bottleneck layer, further reducing information redundancy. The idea behind feature channel grouping processing is to divide the input feature channels into several subgroups, extract one or more groups of features each time, and then perform information interaction through channel shuffling processing. The number of repetitions is consistent with the number of bottleneck layers, thereby promoting feature reuse and reducing the number of parameters and computational complexity. This invention processes the input feature channels in two groups, as follows:

[0060]

[0061]

[0062]

[0063] in, Indicates the number of channels is n Feature map, Representation of feature map The first in n One channel, This represents the bottleneck layer output after feature channel grouping processing. This represents a module consisting of two 3×3 convolutions, an EMA attention mechanism, and residual connections. This indicates that the feature maps are spliced ​​and aggregated along the channel dimension.

[0064] Feature channel grouping processing extracts features from only a subset of channels, preserving original features while reducing parameter count and computational complexity, thus achieving feature reuse. To enhance channel interaction capabilities, channel shuffling is introduced to construct a new feature reuse bottleneck layer, improving effective feature extraction capabilities. This bottleneck layer is embedded into the 3rd and 5th layer C2f architecture to form an EC-C2f structure, constructing a deep feature reuse chain. Through module stacking, continuous preservation and progressive refinement of defect features are achieved, avoiding redundant computation to improve efficiency, and enhancing the model's defect feature representation capabilities and overall detection accuracy and speed.

[0065] Furthermore, a dual-path guided feature complementary fusion module is designed to effectively align shallow location information with deep semantic information. In transmission line defect detection, due to complex backgrounds and large differences in target scale, the spatial location information and semantic information in deep feature maps are easily insufficiently integrated, leading to blurred localization of small targets and feature misalignment. To alleviate this problem, a dual-path guided feature complementary fusion module is designed to replace the C2f structure of the last two layers in the YOLOv8m backbone network. This module consists of two paths: "channel feature guidance" and "spatial feature guidance," which recalibrate the input features respectively. The channel guidance path generates weight vectors through feature encoding to enhance the response to key channels; the spatial guidance path preserves and strengthens positional structure information through convolution operations. The two outputs are fused through 1×1 convolution to form a feature representation with rich spatial details and discriminative semantics. The specific construction of the dual-path guided feature complementary fusion module is as follows: (1) Channel splitting: Assume the input feature map is It is split into two parts along the channel dimension, one part contains The channel, another part contains aisle, The value of is crucial. In deeper layers of the network, branches with low-level spatial information become more prominent. More low-level spatial information is encoded into high-dimensional representations and further transmitted, effectively enhancing the model's ability to perceive the structure of small-scale defects such as insulator damage and bird nests, and improving the collaborative performance of model localization and classification. This stage is described as follows:

[0066] in, .

[0067] (2) Dual-path guidance: In order to obtain spatial mappings of semantic and location information respectively, the following methods are used: The input consists of a position-preserving branch composed of pointwise convolutions (1×1 convolutions). Its receptive field is small, effectively preserving fine-grained spatial structure and positional information from shallow layers within the input features. The output of this branch is represented as... .Will The input consists of a semantic enhancement branch composed of standard 3×3 convolutions, responsible for extracting richer contextual semantic features in the channel dimension. The output of this branch is represented as... The transformation process is shown below:

[0068] in, A function representing the mapping relationship between the learning space and semantic information. It retains a large amount of original spatial location information. It contains rich channel semantic information.

[0069] (3) Complementary Enhancement: To improve the completeness of feature representation and matching accuracy, a complementary enhancement mechanism is proposed to guide the bidirectional fusion of semantic and spatial features. This mechanism uses a spatial feature guidance module to... Each location of important information is assigned a unique weight, which is then mapped to information rich in semantic information. In this process, the structure perception is enhanced; simultaneously, through the channel feature guidance module, unique weights are assigned to important information in each channel, and then these weights are mapped to information with low-level spatial location information. In this process, the semantic hierarchy is enhanced. This process enables strong features to guide weak features, effectively alleviating the imbalance between semantic and spatial information.

[0070] Spatial Feature Guidance Module: First, a 1×1 convolutional layer is used to... The number of channels is reduced to one dimension, then batch normalization is used to obtain spatial information at each location, and finally, a sigmoid activation function layer is used to obtain the key spatial information weights. These weights are then spatially unique. The process is described as follows:

[0071] in, express The feature map after dimensionality reduction processing by a 1×1 convolutional layer.

[0072] Channel feature guidance module: First, use deep convolutional layers to... Each channel undergoes convolution, followed by global average pooling to obtain global information for each channel. Finally, key information weights are obtained through a sigmoid activation function layer. Unique weights are generated for each channel. The process is described as follows:

[0073] in, express The feature map obtained after processing by deep convolutional layers.

[0074] (4) Feature fusion: By multiplying point by point, the feature is fused. Its mapping to This allows it to focus more on channel information; by multiplying point by point... Its mapping to This allows the focus to be placed more on spatial information. Then, the two branches are merged to obtain the output features. It contains the feature of a dual complementary mapping of spatial and semantic relationships. The calculation is as follows:

[0075] in, This indicates element-wise multiplication. This indicates element-wise addition.

[0076] Replacing the last two C2f layers of the backbone network with a dual-path guided feature complementary fusion module helps the model retain and propagate the effective spatial location information extracted by the shallow EC-C2f module to the deeper layers of the network. This effectively alleviates the loss of spatial location information of the defect target during the downsampling process of the backbone network, enhances the model's spatial location perception ability, and improves the model's detection accuracy.

[0077] Furthermore, a cascaded multi-kernel module is constructed to alleviate the feature loss problem of deep small targets in the network, enhance multi-scale perception, and reduce model complexity. Specifically, since defect targets in transmission line images often contain only a few pixels, these targets are prone to feature vanishing during downsampling feature extraction in the backbone network. Therefore, a cascaded multi-kernel module is proposed to fully utilize limited feature information and enhance the network's ability to perceive defect targets at different scales. The cascaded multi-kernel module consists of deep convolutional kernels of different sizes, and spatial point convolutions are added between these kernels to focus on details at different scales, highlighting multi-scale feature representation. The overall structure of the cascaded multi-kernel module is as follows:

[0078] in, These represent the input and output feature maps, respectively. Indicates the kernel size as Depth convolution. This represents pointwise convolution.

[0079] The last downsampling convolutional layer of the backbone network is replaced by a cascaded multi-core module to alleviate the problem of small target feature loss in deep backbone networks and further simplify the network structure, reducing the number of network parameters and computational cost.

[0080] Specifically, the experiment used the PyTorch framework and utilized the GPU. The experimental environment is shown in Table 1, and the basic training parameters are shown in Table 2.

[0081] Table 1

[0082] Table 2

[0083] When drones inspect high-voltage transmission lines and towers, they use their high-definition cameras to capture images of insulators, bird nests, and honeycomb-like foreign objects on the transmission lines. Images that are too poor in quality to be used due to shooting errors such as blurriness, overexposure or underexposure, or poor shooting angle are removed to obtain a preliminary raw sample set.

[0084] Due to the diverse types of defects and foreign objects on high-voltage transmission lines, this invention primarily focuses on detecting insulator damage, insulator flashover, insulator spontaneous explosion, honeycomb defects, and bird nest defects. The LabelImg annotation tool is used to offline annotate various defect features on the original sample set, including insulator spontaneous explosion, bird nest defects, and honeycomb defects. A richer sample set is obtained by fusing the publicly available dataset with the labeled original sample set, encompassing five typical defects: insulator damage, insulator flashover, insulator spontaneous explosion, bird nest defects, and honeycomb defects. The newly constructed sample set is divided into a training set and a test set. Data augmentation operations such as rotation, translation, random cropping, brightness adjustment, and noise addition are introduced into the training set to enrich the image's morphology and scale representation and improve the reproduction frequency of various defects. The final dataset contains 2140 training images and 731 test images.

[0085] Step 4: Train the improved insulator defect and foreign object detection model DD-YOLOv8m on the cloud server using the preprocessed sample dataset to obtain the final detection model. This includes: training the model using the training set, adjusting the hyperparameters based on each training result, including batch size, initial learning rate (lr0), input image size (imgsz), optimizer, etc., until the result tends to be optimal to obtain the final detection model.

[0086] Step 5: Model Deployment Phase. The trained detection model is deployed to a low-altitude inspection drone platform with lightweight computing power for testing. This includes: first, establishing a hardware connection with the drone platform and configuring the runtime environment on its lightweight computing platform, including installing project dependencies, creating a workspace, and deploying the model code repository. Then, a topic node program for image publishing, result subscription, and visualization is written and launched. After system compilation, a simulation test is conducted using a moving camera to verify functionality. After deployment, the trained model is quantitatively evaluated using a test set. Key metrics include precision (P), recall (R), F1 score, mean average precision (mAP), number of parameters (Params), and computational cost (FLOPs) to comprehensively measure its detection performance. The formulas for evaluating model performance are as follows:

[0087]

[0088] in, TP This indicates the number of positive examples that were correctly predicted. TN The number of counterexamples that were correctly predicted; FP This represents the number of instances where the actual value is negative but were incorrectly predicted as positive. FN This represents the number of instances where the true value is positive but was incorrectly predicted as negative.

[0089] Precision measures the model's accuracy in predicting positive classes; recall measures the model's false negative rate, with a higher recall indicating a lower false negative rate, and is one of the important indicators for measuring model performance; mAP measures the model's detection performance across different classes, providing an overall evaluation, with a higher mAP value indicating stronger overall detection capabilities and better performance. Table 3 shows the test results of the DD-YOLOv8m model on the defect dataset, Table 4 shows the ablation experiment results, and Table 5 shows the relevant configuration of the low-altitude inspection UAV platform used.

[0090] Table 3

[0091] Table 4

[0092] “√” indicates that the corresponding module was used in the experiment, and “×” indicates that the corresponding module was not used in the experiment.

[0093] Experiment 1 shows the performance of the unmodified YOLOv8 baseline model. The results show that adding the EMA attention mechanism improves mAP50 by 1.4%, while the increase in parameters and computational cost is negligible, indicating that EMA can improve detection accuracy by focusing on globally important features of insulator defects and foreign objects. Compared to YOLOv8, adding feature channel grouping and shuffling reduces model parameters and computational cost by 6.24M and 19.1G, respectively, while improving mAP50 by 2.1%. This verifies that this strategy reduces model parameters and computational cost while promoting information interaction between feature channels, thereby improving detection accuracy. Replacing the C2f structure with the DG-CFM module improves mAP50 by 1.7%, while reducing parameters and computational cost by 7.8M and 19.1G, respectively. This indicates that the dual-path guided feature complementary fusion module can effectively align shallow positional information with deep semantic information and reduce model redundancy. Replacing the last downsampling convolution with the CMKM module resulted in a slight decrease in detection accuracy, but the model's parameter count decreased by 20.09M, making the model more lightweight and suitable for practical detection tasks. Using both the DG-CFM and CMKM modules simultaneously improved mAP50 by 1.6%, while significantly reducing model parameter count and computational cost, indicating a positive synergistic effect between the two modules. Comparison of experiments 8, 9, and 10 reveals a synergistic effect among the four proposed improved modules. While a single module might negatively impact detection accuracy, their combined use significantly improves accuracy. Compared to the YOLOv8m baseline model, the DD-YOLOv8m model achieved a 3% improvement in mAP50, with parameter count and computational cost decreasing by 24.8M and 16.6G respectively, representing reductions of 76.61% and 19.44%. This further demonstrates the lightweight and effectiveness of the improved algorithm, providing a reliable solution for intelligent inspection of power transmission and distribution networks.

[0094] Table 5

[0095] Example 2: An electronic device, comprising a memory and a processor; The memory is used to store computer programs; The processor is configured to implement the method described in Embodiment 1 when executing the computer program.

[0096] Example 3: A computer-readable storage medium storing a computer program that, when executed by a processor, implements the method described in Embodiment 1.

[0097] Example 4: A computer program product includes a computer program that, when executed by a processor, implements the method described in Example 1.

[0098] In the above embodiments, the reference to "this embodiment" in the specification indicates that a specific feature, structure, or characteristic described in connection with the embodiment is included in at least some embodiments, but not necessarily all embodiments. Multiple appearances of "this embodiment" do not necessarily all refer to the same embodiment.

[0099] In the above embodiments, although the invention has been described in conjunction with specific embodiments thereof, many substitutions, modifications, and variations of these embodiments will be apparent to those skilled in the art from the foregoing description. For example, other memory structures (e.g., dynamic RAM (DRAM)) may be used with the embodiments discussed. The embodiments of the invention are intended to cover all such substitutions, modifications, and variations falling within the broad scope of the appended claims.

[0100] As will be understood by those skilled in the art, the computer-readable storage medium described in this embodiment allows for the implementation of all or part of the steps in the above method embodiments by computer program-related hardware. The aforementioned computer program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the above method embodiments; and the aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks.

[0101] The electronic terminal provided in this embodiment includes a processor, a memory, a transceiver, and a communication interface. The memory and the communication interface are connected to the processor and the transceiver and complete communication between them. The memory is used to store computer programs, the communication interface is used to perform communication, and the processor and the transceiver are used to run the computer programs, so that the electronic terminal performs the steps of the above method.

[0102] In this embodiment, the memory may include random access memory (RAM) and may also include non-volatile memory, such as at least one disk storage device.

[0103] The processors mentioned above can be general-purpose processors, including central processing units (CPUs), network processors (NPs), etc.; they can also be digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0104] This invention can be used in a wide range of general-purpose or special-purpose computing system environments or configurations. Examples include: personal computers, server computers, handheld or portable devices, tablet devices, multiprocessor systems, microprocessor-based systems, set-top boxes, programmable consumer electronics, network PCs, minicomputers, mainframe computers, and distributed computing environments including any of the above systems or devices, etc.

[0105] This invention can be described in the general context of computer-executable instructions, such as program modules, that are executed by a computer. Generally, program modules include routines, programs, objects, components, data structures, etc., that perform a specific task or implement a specific abstract data type. This invention can also be practiced in distributed computing environments where tasks are performed by remote processing devices connected via a communication network. In distributed computing environments, program modules can reside in local and remote computer storage media, including storage devices.

[0106] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A lightweight intelligent inspection method for high-voltage transmission lines based on a drone terminal, characterized in that: Includes the following steps: Step 1: Collect images of insulators, bird nests, and honeycomb-like foreign objects during the transmission line inspection process to form the original sample set; Step 2: Label the defects in the original sample set, preprocess the original sample set, and expand the samples using the public dataset to form a defect dataset; The defect data is divided into training and testing sets. Data augmentation is performed on the training set to expand the training sample size. Step 3: Construct the insulator defect and foreign object detection model DD-YOLOv8m, including: on the basis of the YOLOv8m network, using the multi-scale attention mechanism EMA and the feature channel shuffling enhancement model to integrate and preserve the target space details; designing the dual-path guided feature complementary fusion module DG-CFM to promote the alignment of shallow positional information and deep semantic information; and constructing the cascaded multi-core module CMKM to alleviate the problem of loss of deep small target features in the network, enhance multi-scale perception and reduce model complexity. Step 4: Train and optimize the parameters of the insulator defect and foreign object detection model DD-YOLOv8m using the training set; Step 5: Deploy the trained detection model to an inspection drone platform with edge computing power, test the model using a test set, and apply it in practice.

2. The lightweight intelligent inspection method for high-voltage transmission lines based on a drone terminal according to claim 1, characterized in that: In step 1, when the drone inspects the transmission lines and towers, it takes pictures of insulators, bird nests, and honeycomb-like foreign objects using its onboard camera. Unusable images are removed to obtain a preliminary raw sample set.

3. The lightweight intelligent inspection method for high-voltage transmission lines based on a drone terminal according to claim 1, characterized in that: In step 2, various defect features are labeled offline on the original sample set, including insulator spontaneous explosion, bird nest foreign objects, and honeycomb foreign objects. The defect dataset is obtained by fusing the public dataset with the labeled original sample set. The defect dataset includes five defects: insulator damage, insulator flashover, insulator spontaneous explosion, bird nest foreign objects, and honeycomb foreign objects. The newly constructed defect dataset is divided into a training set and a test set. Data augmentation operations are performed on the training set by introducing rotation, translation, random cropping, brightness adjustment, and noise addition.

4. The lightweight intelligent inspection method for high-voltage transmission lines based on a drone terminal according to claim 1, characterized in that: Step 3 describes the integration and preservation of detailed features in the target space based on the YOLOv8m network using the multi-scale attention mechanism EMA and the feature channel shuffling enhancement model. Specifically, this includes replacing the shallow C2f modules in layers 3 and 5 of the backbone network with EC-C2f modules to construct a deep feature reuse chain. The EC-C2f module consists of the multi-scale attention mechanism EMA, feature channel grouping processing, and feature shuffling processing. The Multi-Scale Attention (EMA) mechanism models long-range dependencies, embeds precise location information, captures cross-scale spatial dependencies through parallel sub-networks, and adaptively fuses feature responses from different receptive fields. The feature channel grouping process divides the input feature channels into several subgroups, and extracts one or more groups of features each time. The feature shuffling process is used to exchange information between grouped features. The number of repetitions is consistent with the number of bottleneck layers, thereby promoting feature reuse and reducing the number of parameters and computational complexity.

5. The lightweight intelligent inspection method for high-voltage transmission lines based on a drone terminal according to claim 4, characterized in that: The multi-scale attention mechanism (EMA) specifically includes: (1) Feature grouping: Assume the input feature map is The average grouping is performed along the channel dimension, defined as follows: (2) Parallel sub-network processing: Attention weights of grouped feature maps are extracted simultaneously for each segment through two parallel sub-networks; in the 1×1 convolution branch, average pooling is performed on the input along the X-axis and average pooling is performed along the Y-axis, and the channels are encoded along the two spatial directions respectively; the 3×3 convolution branch uses a single 3×3 kernel stacked to capture multi-scale feature representations; In the 1×1 convolution branch, the input feature map is first subjected to global average pooling in the horizontal and vertical directions to generate two feature maps with dimensions of C / / G×H×1 and C / / G×1×W, respectively. This obtains long-range dependencies in the horizontal direction and long-range interactions in the spatial direction, while preserving precise positional information in the vertical and horizontal dimensions. in, For the input feature map, These represent the height and width of the input feature map, respectively. The first c Each channel is at a height H The output at the width W The output at that location, For feature map in width W Location and height H Pixel value at; The generated feature map of size C / / G×H×1 is transformed, then concatenated with the feature map of size C / / G×1×W. A shared 1×1 convolution is used for further transformation, followed by channel splitting, resulting in two independent tensors. and : in, These represent two 1×1 convolution transformation functions. This represents the sigmoid activation function; Spatial attention is obtained by multiplying the original grouped input with the obtained attention score; the 3×3 convolution branch captures local cross-channel interactions through 3×3 convolution, expanding the feature space; (3) Cross-spatial learning: The two output tensors of the processed 1×1 convolutional branch and 3×3 convolutional branch are used to encode the global spatial information of the output in the 1×1 convolutional branch using 2D global average pooling and to model the long-range dependency. Before the joint activation mechanism of the channel features, the output of the minimum branch is directly converted into the corresponding dimensional shape, i.e. The expression for 2D global average pooling is as follows: The outputs of parallel processing are multiplied by matrix dot products to generate the first spatial attention map. 2D global average pooling is used to encode the global information in the 3×3 convolutional branches. Before the joint activation mechanism for channel features, the global spatial information of the 1×1 branches is converted into its corresponding dimensional shape. Derive the second spatial attention map; Finally, the output feature maps within each group are aggregated using two generated spatial attention weight values ​​to capture pixel-level pairings and highlight the global context of all pixels. After embedding the multi-scale attention mechanism (EMA) into the second convolution of the bottleneck layer of the C2f module in layers 3 and 5 of the backbone network, the original features are preserved through identity mapping connections. The principle is as follows: in, This represents the output of the bottleneck layer structure after introducing the coordinate attention mechanism. Indicates the input feature map, This represents an integrated module consisting of two 3×3 convolutions and an EMA attention mechanism.

6. The lightweight intelligent inspection method for high-voltage transmission lines based on a drone terminal according to claim 5, characterized in that: The feature channel grouping process divides the input feature channels into two groups for processing. The specific principle is as follows: in, Indicates the number of channels is n Feature map, Representation of feature map The first in n One channel, This represents the bottleneck layer output after feature channel grouping processing. This represents a module consisting of two 3×3 convolutions, a multi-scale attention mechanism (EMA), and residual connections. This indicates that the feature maps are spliced ​​and aggregated along the channel dimension.

7. The lightweight intelligent inspection method for high-voltage transmission lines based on a drone terminal according to claim 4, characterized in that: Step 3 describes designing a dual-path guided feature complementary fusion module (DG-CFM) to promote the alignment of shallow location information with deep semantic information. Specifically, this includes replacing the C2f modules in layers 7 and 9 of the backbone network with the dual-path guided feature complementary fusion module. This module consists of two paths: channel feature guidance and spatial feature guidance, which recalibrate the input features respectively. The specific construction is as follows: (1) Channel splitting: Assume the input feature map is It is split into two parts along the channel dimension, one part contains The channel, another part contains The channel is described as follows: in, ; (2) Dual-path guidance: The input consists of a position-preserving branch composed of pointwise convolutions, which preserves the fine-grained spatial structure and positional information from shallow layers in the input features. The output of this branch is represented as... ;Will The input consists of a semantic enhancement branch composed of standard 3×3 convolutions, responsible for extracting contextual semantic features along the channel dimension. The output of this branch is represented as... The transformation process is represented as follows: in, A function representing the mapping relationship between the learning space and semantic information. The original spatial location information was preserved. It contains semantic information of the multichannel; (3) Complementary enhancement: Through spatial feature guidance modules, for Weights are assigned to important information at each location, and these weights are mapped to... In this process, its structural awareness is enhanced; through the channel feature guidance module, weights are assigned to important information in each channel, and these weights are mapped to... In this context, enhance its semantic hierarchy; (4) Feature fusion: By multiplying point by point, the feature is fused. Its mapping to Above; by multiplying point by point... Its mapping to Then, the two branches are merged together to obtain the output features. It contains the feature of a dual complementary mapping of spatial and semantic relationships. The calculation is as follows: in, This indicates element-wise multiplication. This indicates element-wise addition.

8. The lightweight intelligent inspection method for high-voltage transmission lines based on a drone terminal according to claim 7, characterized in that: The spatial feature guidance module first uses a 1×1 convolutional layer to... The number of channels is reduced to one dimension, and then batch normalization is used to obtain spatial information at each position. Finally, a sigmoid activation function layer is used to obtain the key spatial information weights; unique weights are generated in space. The process is described as follows: in, express The feature map after dimensionality reduction processing by a 1×1 convolutional layer.

9. The lightweight intelligent inspection method for high-voltage transmission lines based on a drone terminal according to claim 7, characterized in that: The channel feature guidance module first uses deep convolutional layers to... Each channel undergoes convolution, followed by global average pooling to obtain global information for each channel. Finally, key information weights are obtained through a sigmoid activation function layer, generating unique weights for each channel. The process is described as follows: in, express The feature map obtained after processing by deep convolutional layers.

10. The lightweight intelligent inspection method for high-voltage transmission lines based on a drone terminal according to claim 4, characterized in that: Step 3, which involves constructing a cascaded multi-kernel module (CMKM), aims to alleviate the problem of feature loss in deep, small targets, enhance multi-scale perception, and reduce model complexity. Specifically, this includes replacing the last downsampling convolution in the backbone network with a cascaded multi-kernel module. This module consists of depthwise convolutional kernels of different sizes, with spatial point convolutions added between these kernels to focus on details at different scales, highlighting multi-scale feature representation. The overall construction of the cascaded multi-kernel module is as follows: in, These represent the input and output feature maps, respectively. Indicates the kernel size as Depth convolution; This represents pointwise convolution.

Images