Power transmission line fault detection method, computer equipment and readable storage medium

Abstract

The invention relates to a power transmission line fault detection method, computer equipment and a readable storage medium. The method comprises the following steps: acquiring an initial image obtained by shooting a power transmission line, and processing the initial image by adopting a space-to-depth convolutional network to obtain a target image; processing the target image by adopting a normalized attention module to obtain an output image; a sub-sampling module is adopted to carry out sampling processing on the output image to obtain a to-be-processed image; and mapping the to-be-processed image to the classification space to generate a fault detection result of the power transmission line. By adopting the method, the accuracy of power transmission line detection can be improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer vision, in particular to a power transmission line fault detection method, a computer device and a readable storage medium. BACKGROUND

[0002] The power transmission line is located in a special geographical location and environmental condition, has many tower points, a long line, a wide coverage, and is exposed in the wild all year round. In addition to being subjected to natural disasters such as strong wind, lightning, and heavy rain, the power transmission line is also often affected by human factors. Abnormal faults such as forest fires, tree barriers, and external damage often occur in long-distance overhead power transmission lines under complex climate and environment. Tripping events caused by external damage occur frequently, which seriously affects the safe operation of the power transmission line.

[0003] In related technologies, in order to effectively respond to the external damage safety hazard of the power transmission line, a video / image monitoring terminal is used to cover the key section of the power transmission line passage, image acquisition of the power transmission line is realized, and a target detection algorithm based on deep learning is used to analyze the acquired image to obtain a fault detection result of the power transmission line.

[0004] However, the physical features of the power transmission line are small, and the target detection algorithm based on deep learning has the problem of insufficient accuracy when applied to power transmission line detection. SUMMARY

[0005] Therefore, the present application provides a power transmission line fault detection method, a computer device and a readable storage medium, which can improve the accuracy of power transmission line detection.

[0006] In a first aspect, the present application provides a power transmission line fault detection method, which includes: acquiring an initial image obtained by photographing a power transmission line, processing the initial image by using a spatial-to-depth convolutional network to obtain a target image; processing the target image by using a normalized attention module to obtain an output image; processing the output image by using a subsampling module to obtain a to-be-processed image; and mapping the to-be-processed image to a classification space to generate a fault detection result of the power transmission line.

[0007] In some embodiments, the processing of the initial image by using the spatial-to-depth convolutional network to obtain the target image includes: dividing the initial image into a plurality of sub-images, and extracting pixel data of the same position from each of the plurality of sub-images; merging the extracted pixel data to obtain a plurality of rearranged images; splicing the plurality of rearranged images in the channel dimension to obtain a dimension spliced image; and performing convolution processing on the dimension spliced image to obtain the target image.

[0008] In some embodiments, the target image is processed by using a normalization attention module to obtain an output image, including: performing normalization processing on each channel image in the target image to obtain normalized image data of each channel, and performing scaling and translation on the normalized image data of each channel to obtain adjusted image data of each channel; performing normalization processing on an initial weight of each channel to obtain a normalized weight of each channel, and multiplying the adjusted image data of each channel and the normalized weight of each channel to obtain an initial channel enhanced image of each channel; performing normalization processing on the initial channel enhanced image of each channel by using an activation function to obtain a channel enhanced feature image; performing normalization processing on a weight of each pixel in the channel enhanced feature image in a depth dimension to obtain a normalized weight of each depth, and multiplying the channel enhanced feature image and the normalized weight of each depth to obtain pixel data of each depth; and performing normalization processing on the pixel data of each depth by using an activation function to obtain the output image.

[0009] In some embodiments, the output image is sampled by using a subsampling module to obtain a to-be-processed image, including: dividing the output image into a plurality of image blocks, and generating an initial graph structure according to the plurality of image blocks; performing feature aggregation and feature updating on the initial graph structure to obtain a graph-level processing graph structure, and fusing the initial graph structure and the graph-level processing graph structure to obtain a fused graph structure; processing the fused graph structure by using a feedforward neural network to obtain a transformed graph structure, and fusing the fused graph structure and the transformed graph structure to obtain a target graph structure; performing sampling processing on the target graph structure to obtain a sampled graph structure, and determining the to-be-processed image according to the sampled graph structure.

[0010] In some embodiments, the initial graph structure is generated according to the plurality of image blocks, including: performing feature extraction on each image block to obtain a feature vector of each image block, and mapping the feature vector of each image block to each node in the graph structure to obtain a node set; calculating the similarity between each node and other nodes, and determining the nodes corresponding to the maximum preset number of similarities as the neighboring nodes of each node; determining the directed edges from the neighboring nodes of each node to each node as an edge set; and combining the node set and the edge set to generate the initial graph structure.

[0011] In some embodiments, the feature aggregation and the feature updating are performed on the initial graph structure to obtain a graph-level processing graph structure, including: determining relative feature vectors between a feature vector of each node in the initial graph structure and feature vectors of adjacent nodes of the each node; extracting a maximum value in each dimension from the relative feature vectors to obtain a maximized feature of the each node; performing feature aggregation on the feature vector of the each node and the maximized aggregated feature of the each node to obtain an aggregated feature of the each node; and updating the aggregated feature of the each node by using an update weight to obtain the graph-level processing graph structure.

[0012] In some embodiments, the sampling processing is performed on the target graph structure to obtain a sampling graph structure, including: obtaining selection degrees of nodes in the target graph structure, and determining a node corresponding to a selection degree greater than a threshold as a key node; determining the key node and all adjacent nodes of the key node as sampling nodes according to the target graph structure; and determining a combination of the sampling nodes and directed edges between the sampling nodes as the sampling graph structure.

[0013] In some embodiments, the image to be processed is mapped to a classification space to generate a fault detection result of the power transmission line, including: performing global average pooling processing on the image to be processed to obtain a global average pooling feature vector; performing convolution processing on the global average pooling feature vector to obtain a convolution feature vector; and mapping the convolution feature vector to the classification space to generate the fault detection result of the power transmission line.

[0014] In a second aspect, a computer device is provided, including a memory and a processor, the memory stores a computer program, and the processor implements the steps of the method of any one of the first aspect when executing the computer program.

[0015] In a third aspect, a computer readable storage medium is provided, which stores a computer program, and the computer program implements the steps of the method of any one of the first aspect when executed by a processor.

[0016] In the technical scheme provided by the embodiments of the present application, the spatial-to-depth convolutional network is used to convert low-resolution image features into higher representation capabilities, solve the detection problem in small target and low-resolution scenes, and realize accurate detection of blurred images and small targets, thereby improving the accuracy of power transmission line detection. In addition, the output image is processed by the subsampling module, the balance between calculation efficiency and memory consumption is solved by reducing the spatial resolution and feature dimension of the input graph, the model inference process is significantly accelerated while the original detection accuracy is preserved, and in addition, the normalization attention module is used to introduce a lightweight channel and spatial attention mechanism to solve the bottleneck problem of insufficient key feature extraction, and effectively enhance the detection performance while keeping the model small and low in computation. BRIEF DESCRIPTION OF DRAWINGS

[0017] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the related art, the drawings needed to be used in the description of the embodiments of the present application or the related art will be briefly introduced. Obviously, the drawings in the following description only constitute some embodiments of the present application, and for those skilled in the art, other related drawings can also be obtained without creative labor on the basis of these drawings.

[0018] Figure 1 The flowchart of the power transmission line fault detection method provided for some embodiments is shown in the figure.

[0019] Figure 2 The flowchart of the method for processing the initial image to obtain the target image by using the space-to-depth convolutional network provided for some embodiments is shown in the figure.

[0020] Figure 3 The process diagram for using the SPD-Conv with a period of 2 is shown in the figure.

[0021] Figure 4 The flowchart of the method for processing the target image to obtain the output image by using the normalized attention module provided for some embodiments is shown in the figure.

[0022] Figure 5 The principle diagram of the channel attention submodule provided for some embodiments is shown in the figure.

[0023] Figure 6 The flowchart of the method for processing the output image to obtain the to-be-processed image by using the subsampling module provided for some embodiments is shown in the figure.

[0024] Figure 7 The diagram of the nodes and edges selected by the height priority subsampling strategy provided for some embodiments is shown in the figure.

[0025] Figure 8 The flowchart of the method for mapping the to-be-processed image to the classification space to generate the fault detection result of the power transmission line provided for some embodiments is shown in the figure.

[0026] Figure 9 The network structure diagram of the SSA-vision Gnn provided for some embodiments is shown in the figure.

[0027] Figure 10 The diagram of the relationship between the precision and the training period of the power transmission line fault detection under the Vision Gnn algorithm and the SSA-Vision Gnn algorithm is shown in the figure.

[0028] Figure 11 The diagram of the relationship between the recall rate and the training period of the power transmission line fault detection under the Vision Gnn algorithm and the SSA-Vision Gnn algorithm is shown in the figure.

[0029] Figure 12 Fig. 4 is a schematic diagram of the relationship between the average accuracy and the training period of the fault detection of the transmission line under the Vision Gnn algorithm and the SSA-Vision Gnn algorithm when the cross-correlation ratio threshold is 0.5;

[0030] Figure 13 Fig. 5 is a schematic diagram of the structure of a computer device provided by some embodiments. DETAILED DESCRIPTION

[0031] The embodiments of the technical solutions of the present application will be described in detail below with reference to the drawings. The following embodiments are only used to more clearly illustrate the technical solutions of the present application, and therefore only serve as examples, and cannot limit the protection scope of the present application.

[0032] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs; the terms used herein are only for the purpose of describing specific embodiments and are not intended to limit the present application; the terms "include" and "have" and any variations thereof in the specification and claims of the present application and the above description of drawings are intended to cover non-exclusive inclusion.

[0033] In the description of the embodiments of the present application, the technical terms "first", "second", etc. are only used to distinguish different objects, and cannot be understood as indicating or implying relative importance or implicitly indicating the number, specific order or primary and secondary relationship of the indicated technical features. In the description of the embodiments of the present application, the meaning of "a plurality of" is two or more, unless otherwise explicitly specified. In the description of the embodiments of the present application, "each" means each or each of a plurality, unless otherwise explicitly specified.

[0034] Reference herein to "an embodiment" means that a particular feature, structure, or characteristic described in connection with the embodiment can be included in at least one embodiment of the present application. The appearance of the phrase in various places in the specification does not necessarily all refer to the same embodiment, nor is it necessarily independent or alternative embodiments to other embodiments. It is explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0035] In the description of the embodiments of the present application, the term "and / or" is only a description of the association relationship of the associated objects, which means that there can be three relationships, for example, A and / or B, which can represent the three cases of A alone, A and B together, and B alone. In addition, the character " / " in this paper generally represents that the front and rear associated objects are a "or" relationship.

[0036] In order to effectively deal with the safety hidden danger of external damage of power transmission line construction, the key section of the power transmission line channel is covered through the video / image monitoring terminal, and the remote online patrol of the safety hidden danger of the power transmission line channel is realized. Through the video online patrol, the effective prevention of important line construction and other safety hidden dangers is realized, and the operation and maintenance efficiency of the power transmission line is improved.

[0037] Traditional power transmission channel hazard detection mainly uses target detection algorithm, which detects the edge, color and other features of the image, such as background modeling and straight line detection algorithm based on video monitoring system, tracks the external invasion and automatically warns, but this method has single criterion index, which is easy to cause false negatives. Laser radar technology acquires power transmission line scene modeling, calculates the spatial coordinates of the line and the object, and judges whether the object threatens the line, but this method has high cost and needs additional software and hardware. At present, machine learning technologies such as deep learning have made great progress, and intelligent identification of power transmission line construction machinery, large vehicles and other external damage using deep learning technology can greatly improve the accuracy and real-time performance of fault prediction, improve work efficiency and save cost.

[0038] In today's society, electricity is a key energy to support economic development and people's daily life, and its stable supply is crucial. Timely and accurate detection of power transmission line faults and rapid adoption of effective repair measures are of great significance to ensure the reliable operation of the power system, reduce power loss, and improve power supply quality. Traditional power transmission line fault detection is mainly based on manual inspection, which requires work in the suburbs or wild areas and manual tower climbing, and the power transmission line fault is judged by experience with the naked eye. Therefore, manual inspection has the disadvantages of insecurity, high cost, low precision, and low detection rate. Compared with traditional detection methods, using unmanned aerial vehicle to take pictures of power transmission lines and then using deep learning-based power transmission line detection technology can realize safe, low-cost, high-precision, and high-detection-rate power transmission line fault detection in different backgrounds. In the aspect of power transmission line detection, the convolutional neural network (CNN) automatically learns the deep features of the image layer by layer, and through large-scale training to optimize the network model parameters, it has strong feature extraction ability and generalization ability.

[0039] However, the deep learning-based power line detection scheme in the related art has poor detection effect in a complex background. The power line is densely arranged in various complex environments, so the physical features of each detection target of the power line are relatively small, and the existing power line target detection is prone to lose the detection target. In addition, the small features of the power line detection target are often in a complex environment (such as a cluttered background, complex occlusion, and small targets), which further causes the extraction of local features to be affected by local noise. The current deep learning-based target detection algorithm applied to power line detection has the problem of insufficient accuracy. In addition, the loss of detailed features caused by the convolution layer commonly used in neural networks will cause the problem of unsatisfactory detection effect of low-resolution images and small targets not being normally detected.

[0040] In view of the problems of insufficient accuracy in the current various deep learning-based power line detection technologies in a complex background, and the problems of unsatisfactory detection effect of low-resolution images and small targets not being normally detected in the detailed detection tasks such as low-resolution image detection and small target detection, an embodiment of the present application proposes a visual feature modeling method based on graph convolution. The graph neural network (Gnn) theory is used to dynamically construct local node relationship and global information fusion mechanism, and a graph neural network model combining graph convolution and convolution operation is established.

[0041] Figure 1 A flowchart of a power line fault detection method provided for some embodiments is shown in FIG. 1, which is applied to a computer device. The method comprises the following steps. Figure 1

[0042] S101, an initial image obtained by photographing a power line is acquired, and a space-to-depth convolution (SPD-Conv) network is used to process the initial image to obtain a target image.

[0043] In some embodiments, the image photographing of the power line can be performed by using a drone or a fixedly installed camera.

[0044] In some embodiments, the initial image can be an original image obtained by photographing. In other embodiments, the initial image can be obtained by preprocessing the original image obtained by photographing. Illustratively, the image preprocessing can include at least one of the following: rotation, scaling, enlargement, denoising, etc.

[0045] ​SPD-Conv is a technique that converts image spatial information into depth information, enabling the convolutional neural network to learn image features more effectively. This method optimizes the network's processing capability for small objects and low-resolution images by reducing information loss and improving feature extraction accuracy.

[0046] S102, processing the target image by using a normalization-based attention module (NAM) to obtain an output image.

[0047] In some embodiments, the normalization-based attention module includes a channel attention submodule and / or a spatial attention submodule.

[0048] In some embodiments, S102 can include processing the target image by using a channel attention submodule to obtain an output image.

[0049] In other embodiments, S102 can include processing the target image by using a spatial attention submodule to obtain an output image.

[0050] In yet other embodiments, S102 can include processing the target image by using a channel attention submodule to obtain a channel-enhanced feature image; and processing the channel-enhanced feature image by using a spatial attention submodule to obtain an output image.

[0051] S103, processing the output image by using a subsampling module to obtain a to-be-processed image.

[0052] In some embodiments, the output image is divided into multiple image blocks (e.g., N image blocks), and the subsampling module is used to process the multiple image blocks to obtain M image blocks, where M is less than N. The M image blocks are determined as the to-be-processed image.

[0053] Wherein, processing the output image by using the subsampling module includes the following steps: graph structure generation (or construction), feature aggregation and multi-head update, and height priority subsampling strategy.

[0054] S104, mapping the to-be-processed image to a classification space to generate a fault detection result of the power transmission line.

[0055] In some embodiments, S104 can include performing a global average pooling on the to-be-processed image, then performing a convolution operation, mapping to a classification space, and generating a fault detection result of the power transmission line through the classification space.

[0056] ​Exemplarily, after global average pooling is performed on the image to be processed, a convolution operation is performed to obtain a convolution feature vector, and the convolution feature vector is input to a fully connected layer. The fully connected layer performs further linear transformation and non-linear mapping on the feature, and maps it to the final classification dimension. An activation function such as a Softmax function or a Sigmoid function (which is suitable for binary classification) is applied to the output of the fully connected layer. The Softmax function converts the output vector into a probability distribution of each class, and the Sigmoid function maps the output value to between 0 and 1, and the class with the largest probability value can be selected as the fault detection result according to a threshold.

[0057] Exemplarily, after global average pooling is performed on the image to be processed, a convolution operation is performed to obtain a convolution feature vector, and the convolution feature vector is input to a fully connected layer. The fully connected layer performs further linear transformation and non-linear mapping on the feature, and maps it to the final classification dimension. An activation function such as a Softmax function or a Sigmoid function (which is suitable for binary classification) is applied to the output of the fully connected layer. The Softmax function converts the output vector into a probability distribution of each class, and the Sigmoid function maps the output value to between 0 and 1, and the class with the largest probability value can be selected as the fault detection result according to a threshold.

[0058] Exemplarily, the fault detection result of the power transmission line can include a result of a fault of the power transmission line or a result of normality of the power transmission line.

[0059] In the technical scheme provided by the embodiments of the present application, a spatial-to-depth convolution (SPD-Conv) network is used to convert low-resolution image features into higher representation capabilities, solve the detection problem in small target and low-resolution scenes, and realize accurate detection of fuzzy images and small targets, thereby improving the accuracy of power transmission line detection. In addition, a subsampling module is used to process the output image, by reducing the spatial resolution and feature dimension of the input image, a balance between calculation efficiency and memory consumption is solved, the original detection accuracy is preserved while the model inference process is significantly accelerated, and in addition, a normalization attention module is used to introduce a lightweight channel and spatial attention mechanism to solve the bottleneck problem of insufficient key feature extraction, and effectively enhance the detection performance while keeping the model small and low in computation.

[0060] Next, the step of using a spatial-to-depth convolution network to process the initial image in S101 to obtain a target image is further described: Figure 2 For the method of using a spatial-to-depth convolution network to process the initial image to obtain a target image provided by some embodiments, a flowchart is shown in Figure 2 As shown in the figure, the method is applied to a computer device, and the method comprises the following steps:

[0061] S1011, divide the initial image into a plurality of sub-images, and extract pixel data of the same position from each of the plurality of sub-images.

[0062] The size of each sub-image can be 2x2, 3x3 or 4x4. The number of channels of each sub-image is the same as the number of channels of the initial image.

[0063] For example, in the case of a sub-image of 2x2, the upper left pixel data in each 2x2 image is extracted, the upper right pixel data in each 2x2 image is extracted, the lower left pixel data in each 2x2 image is extracted, and the lower right pixel data in each 2x2 image is extracted.

[0064] S1012, merge the extracted pixel data to obtain a plurality of rearranged images.

[0065] Exemplarily, the pixel data at the same position of different channels can be merged to obtain a plurality of rearranged images. For example, the upper left pixel data is merged to obtain a first rearranged image, the upper right pixel data is merged to obtain a second rearranged image, the lower left pixel data is merged to obtain a third rearranged image, and the lower right pixel data is merged to obtain a fourth rearranged image.

[0066] S1013, the plurality of rearranged images are spliced in the channel dimension to obtain a dimension spliced image.

[0067] For example, the first rearranged image, the second rearranged image, the third rearranged image, and the fourth rearranged image are spliced in the channel dimension to obtain a dimension spliced image.

[0068] S1014, the dimension spliced image is subjected to convolution processing to obtain a target image.

[0069] Exemplarily, the dimension spliced image is subjected to convolution processing to obtain a target image with the same depth as the initial image.

[0070] In image detection tasks, when detecting low-resolution, blurred images or small-size targets, the performance of the convolutional neural network used will often decrease, which is caused by the convolution and pooling operations in CNN. CNN extracts high-level features through convolution and pooling, but these operations gradually reduce the resolution of the feature map, making it easy to lose or blur the details of low-resolution images and small targets. Due to the pooling operation, the proportion of small targets on the feature map is further reduced, and after several layers of processing, only a small number of pixels may be left, making it difficult for CNN to effectively identify their features. The size of the receptive field also affects the detection effect. The receptive field of CNN determines the area it focuses on, and small targets are often difficult to match a larger receptive field, resulting in details being ignored.

[0071] The details loss problem of CNN network is mainly caused by the convolution and pooling to extract high-level features, and the convolution and pooling operations are also used in the Vision Gnn model, so the spatial-to-depth convolution (SPD-Conv) is introduced into the model to enhance the network's ability to identify details. SPD-Conv is a convolution method that converts spatial information of input images into depth information. Traditional convolution operations usually extract features by sliding the convolution kernel in the spatial dimension of the image when processing feature maps, while SPD-Conv rearranges the spatial dimension of the input to combine adjacent pixels together, thereby converting spatial information into depth information. The SPD-Conv module includes a space-to-depth (SPD) layer and a convolution layer with a step size of 1.

[0072] The following is the process of using SPD-Conv to replace the step convolution layer and the pooling layer:

[0073] The space-to-depth SPD layer replaces the convolution layer: the initial feature map has a size of (32x32) and contains detailed information of small objects. After passing through the SPD layer, the spatial resolution is reduced to (16x16), and the reduced spatial information is transferred to the channel dimension, so that the number of channels increases without information loss.

[0074] The non-step convolution layer replaces the pooling layer: after SPD layer processing, the size of the feature map becomes (16x16), but the number of channels increases, from the original 64 channels to 256 channels. The non-step convolution layer is applied to the feature map with increased channels, which does not reduce the spatial size of the feature map, but extracts important features by learning the information in the increased channels.

[0075] In this way, even at a smaller spatial resolution, the model can effectively capture the detailed information of small objects.

[0076] Figure 3 The process diagram for using SPD-Conv with a period of 2 is shown in Figure 3 The period of 2 means that the pixels are divided by 2, and there are s / 2 periods in the x direction, and the same in the y direction, so the original feature layer is divided into s 2 / 4 feature maps.

[0077] In Figure 3 (a) shows a standard feature map, i.e. the initial image described above, with a channel number C, a width W and a height H; (b) shows the space-to-depth operation, in which the spatial information is rearranged into depth channels; (c) shows the increase in depth of the resulting feature map; (d) shows the output feature map after convolution with a step size of 1 (i.e. the target image described above).

[0078] Next, the processing of the target image in the step of S102 using the normalized attention module to obtain the output image is further described: Figure 4 For some embodiments, a flowchart of a method for processing a target image using a normalized attention module to obtain an output image is shown in FIG. 1, and the method is applied to a computer device. The method comprises: Figure 4

[0079] S1021, normalizing each channel of the target image to obtain normalized image data of each channel, and scaling and translating the normalized image data of each channel to obtain adjusted image data of each channel.

[0080] S1022, normalizing the initial weight of each channel to obtain normalized weight of each channel, and multiplying the adjusted image data of each channel and the normalized weight of each channel to obtain an initial channel enhanced image of each channel.

[0081] S1023, normalizing the initial channel enhanced image of each channel using an activation function to obtain a channel enhanced feature image.

[0082] S1024, normalizing the weight of each pixel in the channel enhanced feature image in the depth dimension to obtain normalized weight of each depth, and multiplying the channel enhanced feature image and the normalized weight of each depth to obtain pixel data of each depth.

[0083] S1025, normalizing the pixel data of each depth using an activation function to obtain an output image.

[0084] The implementation method of the normalized attention module (NAM) is described as follows:

[0085] NAM is a lightweight and efficient attention mechanism. Its design goal is to improve the performance of neural networks by selectively enhancing important features. Unlike traditional attention mechanisms, NAM does not rely on additional fully connected layers or convolutional layers, but uses the scaling factor of normalization (such as batch normalization) to represent the importance of features. This approach significantly reduces computational complexity and memory requirements, making NAM particularly suitable for lightweight architectures and resource-constrained applications.

[0086] ​NAM has unique advantages compared to other attention mechanisms, as it focuses on the importance of weights based on feature variance. NAM uses standard deviation to measure the significance of weights, thereby avoiding redundant structures in methods such as Squeeze-and-Excitation Network (SENet), Bottleneck Attention Module (BAM), and Convolutional Block Attention Module (CBAM). This feature allows NAM to significantly reduce computational overhead while ensuring effective attention mechanisms, which is particularly beneficial for high-dimensional data models or scenarios with limited computing resources.

[0087] NAM can be seamlessly integrated into various neural network architectures, such as CNN and Residual Neural Network (Resnet). Its lightweight and flexible design is also suitable for deployment in Gnn, enhancing the selectivity of Gnn to important node and edge features, and improving the capture of graph structure information. In Gnn, NAM can help the model focus on weights that contribute significantly to the learning process, thereby improving the overall quality of graph representation.

[0088] NAM combines channel attention and spatial attention sub-modules, using the scaling factor of batch normalization to measure the importance of channels and pixels, thereby achieving effective identification and utilization of features.

[0089] Figure 5 The principle diagram of the channel attention sub-module provided for some embodiments uses the scaling factor of batch normalization (BN) to measure the importance of channels, and obtains output features by calculating weights. For example, the input feature is denoted as F1, with dimensions HxWxC, where H is the height, W is the width, and C is the number of channels. The expression of the batch normalization operation is as follows: ; wherein is the input feature of channel i, i.e., the pixel data in the i-th channel, and are the mean and variance of channel i, respectively, is a very small constant to increase numerical stability, denotes the normalized data of the input feature of channel i. In this way, each pixel data in the i-th channel is subjected to batch normalization operation to obtain the normalized data of each pixel in the i-th channel, i.e., the normalized image data of the i-th channel.

[0090] Then, scaling and translation operations are performed on the normalized data to restore the expression capability of the model, where, and are learnable parameters for adjusting the normalized data: ; denotes the data after scaling and shifting. The image after scaling and shifting the normalized data is the adjusted image data.

[0091] After batch normalization, the normalization weight of each channel is calculated, and the normalization weight of the i-th channel is determined by the following formula: wherein, wherein is the initial weight of the i-th channel. Then the input feature after normalizing the input feature is multiplied by the weight to obtain the intermediate feature value . .

[0092] Then, the feature is normalized using the Sigmoid activation function to ensure that the feature value is between 0 and 1, and the output feature (i.e., the channel-enhanced feature image mentioned above) is obtained; ; ; .

[0093] The output feature of the channel attention sub-module can be calculated as follows: ; wherein, denotes the weight of each channel, denotes the image after performing the batch normalization operation in each channel.

[0094] The spatial attention sub-module applies pixel normalization to measure the importance of pixels to obtain the output feature. The specific calculation process is similar to that of the channel attention sub-module. The output feature of the spatial attention sub-module can be calculated as follows: ; wherein, denotes the output feature of the spatial attention sub-module, denotes the weight of each depth, denotes the output feature of the channel attention sub-module .

[0095] In addition, in some embodiments, the normalization attention module (NAM) also includes a weight sparsity penalty mechanism. This mechanism suppresses less significant weights by adding a regularization term to the loss function, thereby improving the generalization ability and efficiency of the model. x represents the input, y is the output, and W represents the network weight, i.e. and the union of and, l(.) is a loss function, g(.) is a norm penalty function of l1, p is a penalty that balances g(y) and g(A), and the loss function formula after adding the weight coefficient penalty mechanism is as follows: .

[0096] Next, the step of S103 is further described: Figure 6 The flowchart of the method for sampling and processing the output image by using the sub-sampling module to obtain the to-be-processed image provided for some embodiments is shown in FIG. 8, and the method is applied to a computer device. The method includes the following steps. Figure 6

[0097] S1031, the output image is divided into a plurality of image blocks, and an initial graph structure is generated according to the plurality of image blocks.

[0098] In some embodiments, generating the initial graph structure according to the plurality of image blocks can include: performing feature extraction on each image block to obtain a feature vector of each image block, and mapping the feature vector of each image block to each node in the graph structure to obtain a node set; calculating the similarity between each node and other nodes, and determining the nodes corresponding to the maximum preset number of similarities as the neighboring nodes of each node; determining the directed edges from the neighboring nodes of each node to each node as an edge set; and combining the node set and the edge set to generate the initial graph structure.

[0099] The following describes the process of generating (or constructing) the graph structure:

[0100] A graph is a set of nodes and the relationships (i.e., edges) between these nodes. Constructing a graph structure is to construct the node set and the edge set of data. For example, for an input power line image with a size of , first, the image is preprocessed and divided into image blocks of equal size (for example, the image is divided into a grid of , and each block has a size of , obtaining blocks). Then, a feature vector is extracted for each image block, and the is the feature dimension. The feature extraction can use a pre-trained deep neural network or a pre-set feature extraction method (for example, using convolution operation or other operations), and finally form a feature matrix . Each feature vector (the i-th feature vector) is mapped to a node in the graph, obtaining a node set . Then, for each node , its is found according to the similarity of the feature vectors (such as Euclidean distance or cosine similarity).​ is an integer greater than or equal to 1, for example The value of K can be 1, 2, 3, 5, 10, or 20, etc. K nearest neighbor nodes and each neighbor node is added with a directed edge from to , thereby constructing an edge set . Finally, the graph is generated by combining the node set and the edge set , and the construction process of the graph can be represented as . Through this process, the graph formed can capture the spatial distribution and local relationship of the image block features, providing input data for subsequent graph-based algorithms. Among them, the graph is the initial graph structure generated.

[0101] In the embodiments of the present application, the maximum preset number of nodes corresponding to the similarity are determined as the neighboring nodes of each node; the directed edges from the neighboring nodes of each node to each node are determined as the edge set; and the initial graph structure is generated by combining the node set and the edge set. This can specifically select a part of nodes that are most similar and have the highest correlation degree to each node as its neighboring nodes, thereby highlighting the close relationship between nodes, focusing on those connections that are important in the similarity level, avoiding the inclusion of too many irrelevant or weakly related connections when constructing the graph structure, making the generated graph structure more compact and meaningful, and better reflecting the potential similarity patterns and key relationships in the data.

[0102] In some embodiments, the initial graph structure is subjected to feature aggregation and feature updating to obtain a graph-level processing graph structure, and the initial graph structure and the graph-level processing graph structure are fused to obtain a fused graph structure.

[0103] In some embodiments, the initial graph structure is subjected to feature aggregation and feature updating to obtain a graph-level processing graph structure, which can include: determining each relative feature vector between the feature vector of each node in the initial graph structure and the feature vector of each neighboring node of each node; extracting the maximum value in each dimension from each relative feature vector to obtain the maximized feature of each node; performing feature aggregation on the feature vector of each node and the maximized aggregated feature of each node to obtain the aggregated feature of each node; and updating the aggregated feature of each node using an update weight to obtain the graph-level processing graph structure.

[0104] In some embodiments, updating the aggregated features of each node by employing the update weights to obtain the graph-level processed graph structure can include: splitting the aggregated features of each node into h head features, employing h update weights to respectively perform multi-head updating on the h head features of each node to obtain the graph-level processed graph structure.

[0105] In some embodiments, fusing the initial graph structure and the graph-level processed graph structure to obtain the fused graph structure can include: fusing (e.g., adding, multiplying, weighted summing, or weighted averaging) the feature vector of each node in the initial graph structure with the feature vector of the corresponding each node in the graph-level processed graph structure to obtain the fused graph structure.

[0106] The following describes the process of feature aggregation and multi-head updating:

[0107] Matrix features Cover samples, and each sample has dimensional features, thereby constructing feature . Further, based on the features, a graph structure is constructed, and the association between nodes is determined by means of adjacency relations. In the graph convolution layer, the exchange of information between nodes is realized by aggregating and updating the features of adjacent nodes, and the mathematical expression is: .

[0108] wherein, is an updated graph structure obtained by updating . is an update function of the feature matrix, is an aggregation function of the feature matrix, and correspond to the learnable weights required for the aggregation and updating operations, respectively.

[0109] For each feature vector , the updated feature vector of the feature vector is obtained in the following manner: ; wherein, denotes the updated feature vector obtained by updating the feature vector , is a set of neighbor nodes of , denotes an aggregation function of the feature vector, denotes an update function of the feature vector.

[0110] Exemplarily, the aggregation is implemented by using the maximum relative graph convolution, and the model bias is ignored because the model bias has little effect on the model performance. When the updated weight is used to update the aggregated feature of each node, the bias is not considered. Wherein, , wherein, represents , and , and represents the maximum one of the lengths between and the maximum one of the feature vectors,

[0111] Exemplarily, the aggregation operation obtains the representation of the node by aggregating the features of the adjacent nodes. For example, summation, average, maximum, weighted average or weighted summation, etc. are used to obtain the aggregation result of two feature vectors, wherein the weight used in the weighted average or weighted summation can be , wherein the feature vector corresponds to the weight of the maximum one of the feature vectors. For another example, the aggregation result of two feature vectors is obtained by concatenating the two feature vectors, and by concatenating the current node feature and the relative feature change of the neighbor node, the local representation ability of the node can be enhanced.

[0112] The update operation further integrates the aggregated features. The updated features not only retain the original information, but also combine the context information of the neighbor nodes, so that the representation of each node is more comprehensive. Wherein, .

[0113] In some embodiments, in order to enhance the diversity of the features, a multi-head update operation is introduced. Through parallel calculation of the multi-head update, different subspaces can capture multi-perspective information of the features, and improve the expression ability and generalization of the model. The multi-head update operation splits the aggregated feature into , that is, , and then uses the weight (i.e. h update weights) to update the corresponding . All can be updated in parallel and connected as the final value:

[0114] .

[0115] In some embodiments, the feature vector of each node of the initial graph structure is replaced by the result (i.e. the updated feature vector of each node) of each node to obtain a graph-level processed graph structure.

[0116] ​In some embodiments, the feature vector of each node is updated by using a nonlinear activation function to obtain an enhanced feature vector of each node, and the feature vector of each node of the initial graph structure is replaced by the enhanced feature vector of each node to obtain the graph-level processed graph structure.

[0117] In the Deep Graph Convolution Network (Deep GCN) architecture, there is an over-smoothing problem, which weakens the uniqueness of node features and causes the visual recognition performance to decline. In order to alleviate this problem, more feature transformations and nonlinear activation functions are introduced into the backbone network. After each layer of graph convolution operation, a nonlinear activation function is used to enhance the nonlinear expression ability of the features, such as ReLU or Gaussian Error Linear Units (GeLU) and the like. Exemplarily, the ReLU is used in the embodiments of the present application to solve the gradient disappearance problem of the neural network. The feature transformation is realized through a fully connected layer, and the formula is as follows: ; wherein, , and is the weight parameter of the fully connected layer, is the activation function, is the feature after aggregation.

[0118] In some embodiments, the image can be processed by multi-layer feature aggregation and multi-head update, and the output of the feature aggregation and multi-head update of the previous layer is the input of the feature aggregation and multi-head update of the current layer. For example, there can be 3 layers.

[0119] By stacking multiple layers of feature transformation and nonlinear activation, the complex feature interaction relationship is further captured: the formula of the multi-layer stacking is as follows: . Wherein, represents the layer, , is the weight parameter of the fully connected layer in the layer. By changing the feature dimension layer by layer, the feature representation ability is gradually enhanced.

[0120] ​In the technical scheme provided by the embodiments of the present application, the feature vectors of each node are aggregated with the maximum features, and the features of the node itself and the key features of the difference between the node and the adjacent nodes are comprehensively aggregated. The aggregation operation makes the feature representation of the node more rich, and can contain more information about the node itself and the relationship between the node and the surrounding nodes. The rich feature expression capability helps the model better understand the role and status of the node in the graph, improves the performance of the model in processing the graph structure data, for example, the nodes can be more accurately classified in the node classification task, and the node features in the graph structure are optimized and updated, which can more accurately reflect the relationship between the nodes and the overall structure information of the graph. The high-quality graph structure provides a good foundation for subsequent graph classification tasks, and can improve the accuracy and efficiency of these tasks. At the same time, the graph-level processing of the graph structure can also be used as an intermediate result to provide more valuable input for further feature extraction, model training, etc.

[0121] S1033, the fusion graph structure is processed by using a feed-forward neural network (FFN) to obtain a transformed graph structure, and the fusion graph structure and the transformed graph structure are fused to obtain a target graph structure.

[0122] In order to further improve the feature conversion capability and alleviate the over-smoothing phenomenon, the FFN is used on each node. The FFN is a simple multi-layer perceptron, and the specific formula is as follows: . Wherein , and are the weights of the fully connected layer, and are the bias terms. The hidden layer dimension of the FFN is greater than the input dimension, and contains two fully connected layers. In the graph operation GP and the FNN, after each fully connected layer or graph convolution layer operation is completed, a batch normalization operation is performed to balance the feature distribution. Based on the image-based graph representation and the proposed graph convolution block (GCB), a Gnn network architecture suitable for visual tasks can be constructed.

[0123] In some embodiments, the fusion of the fusion graph structure and the transformed graph structure to obtain the target graph structure can include: fusing (for example, adding, multiplying, weighted summing or weighted averaging) the feature vector of each node in the fusion graph structure with the feature vector of each corresponding node in the transformed graph structure to obtain the fusion graph structure.

[0124] S1034, the target graph structure is sampled to obtain a sampled graph structure, and the sampled graph structure is used to determine a to-be-processed image.

[0125] In some embodiments, the sampling processing on the target graph structure to obtain a sampled graph structure can include: obtaining a selection degree of each node in the target graph structure, determining a node corresponding to a selection degree greater than a threshold value as a key node; determining, according to the target graph structure, the key node and all adjacent nodes of the key node as sampled nodes; and determining, as the sampled graph structure, a combination of the sampled nodes and directed edges between the sampled nodes.

[0126] Exemplarily, the selection degree of each node can be obtained by: obtaining a similarity (such as Euclidean distance or cosine similarity) between a feature vector of each node and each adjacent vector, and determining, as the selection degree of each node, a number of similarities greater than a preset value among all similarities corresponding to each node.

[0127] Exemplarily, the selection degree of each node can be obtained by: obtaining a similarity (such as Euclidean distance or cosine similarity) between a feature vector of each node and each adjacent vector, and determining, as the selection degree of each node, an average value, a maximum value, a minimum value, a weighted average value or a weighted summation value among all similarities corresponding to each node.

[0128] Embodiments of the present application introduce a highly preferential sub-sampling strategy. Nodes with higher selection degrees are preferentially selected for sampling, because these nodes are often key nodes (such as central nodes or hub nodes) in the network. By introducing node sub-sampling in the graph construction stage, the nodes or edges with higher degrees in the graph are preferentially selected, the core structure and key nodes of the graph are preserved, and the computational complexity is reduced.

[0129] In some embodiments, determining the image to be processed according to the sampled graph structure can include: determining, as the image to be processed, each image block corresponding to each node in the sampled graph structure. In some embodiments, the image to be processed can be obtained by splicing each image block in the channel dimension. In other embodiments, the image to be processed can be obtained by splicing each image block in the length or width direction.

[0130] Figure 7 A schematic diagram of the nodes and edges selected by the highly preferential sub-sampling strategy provided by some embodiments is shown in FIG. 1. Figure 7 As shown in FIG. 1, the black nodes are nodes with degrees higher than a threshold value, the selected nodes are nodes with higher degrees, the selected nodes are a combination of a black node and a node adjacent to the black node, and the selected edges are edges of the selected nodes.

[0131] Embodiments of the present application use a degree-based sub-sampling method to only retain key nodes with higher connection degrees, which is represented as follows: ; wherein, is the degree of node , and is a threshold value (or a sampling threshold value).

[0132] Figure 8 A flowchart of a method for mapping an image to be processed to a classification space and generating a fault detection result of a power transmission line is provided for some embodiments, as shown in Figure 8 The method is applied to a computer device, Figure 8 Compared with Figure 1 The difference between the embodiments is that S104 includes the following steps:

[0133] S1041, performing global average pooling processing on the image to be processed to obtain a global average pooling feature vector.

[0134] Global average pooling is a pooling operation on the spatial dimension (height and width dimension for images) of the feature map. Unlike ordinary pooling operations (such as max pooling, average pooling, etc., which are usually performed in a local area), global average pooling calculates the average value of the entire feature map. Specifically, for a feature map of size HxWxC, the average value of all pixel values in each channel is calculated separately.

[0135] S1042, performing convolution processing on the global average pooling feature vector to obtain a convolution feature vector.

[0136] S1043, mapping the convolution feature vector to a classification space to generate a fault detection result of a power transmission line.

[0137] In some embodiments, mapping the convolution feature vector to the classification space can include: inputting the convolution feature vector to a fully connected layer, and applying an activation function on the output of the fully connected layer to obtain a probability distribution of each class.

[0138] The network model provided by the embodiments of the present application is an SSA-vision Gnn that integrates Space-to-depth Convolution (SPD-Conv), Subsampling, and Attention.

[0139] Figure 9A schematic diagram of a network structure of the SSA-vision Gnn provided for some embodiments includes: 1) an initial layer; 2) a depth map convolutional network; and 3) an output layer. The initial layer is configured to receive an initial image of an input layer, and perform a spatial-to-depth convolution operation on the initial image to obtain a target image. The target image is subjected to batch normalization to obtain a designated image, and the designated image is subjected to a rectified linear unit to obtain an output image. In the depth map convolutional network, the output image is first constructed into an initial graph structure, and then subjected to graph-level processing, i.e., one or more layers of feature aggregation and multi-head update processing, to obtain a graph-level processing graph structure. The initial graph structure and the graph-level processing graph structure are fused to obtain a fused graph structure. The fused graph structure is then processed by a feedforward neural network to obtain a transformed graph structure, and the fused graph structure and the transformed graph structure are fused to obtain a target graph structure. The target graph structure is sampled to obtain a sampled graph structure, and the sampled graph structure is used to determine the image to be processed. The image to be processed enters the output layer through a classification head. The features output by the depth map convolutional network are subjected to global average pooling and convolution operations, and are mapped to a classification space to generate a final prediction result.

[0140] The initial layer performs preliminary feature extraction on the image without loss of details. The initial layer extracts features of the image through a series of spatial-to-depth convolution (SPD-Conv), batch normalization (BN), and rectified linear unit (ReLU). For example, for an image with a size of 224x224x3, the image size is reduced to 14x14 to achieve downsampling, and the image dimension is increased from 3 to 768 to project the input image from 3 channels to a high-dimensional feature space to achieve feature extraction. The batch normalization (BN) layer and the rectified linear unit (ReLU) activation function between the spatial-to-depth convolution (SPD-Conv) layers are used to enhance the feature expression capability. Subsequently, position embedding is performed. After the convolutional layer and the graph convolutional layer, the positional relationship may be weakened. Therefore, explicit embedding position coding is used to supplement spatial position information and ensure that the network is sensitive to the spatial arrangement of the input image.

[0141] Deep GCN is the core of SSA-vision Gnn, which is composed of multiple stacked graph convolutional blocks (GCBs). Each block contains a graph processing (GP) and a feed-forward neural network (FFN). The GP builds a graph structure through dynamic K-nearest neighbors, capturing local node relationships and gradually fusing global features. The FFN module further optimizes feature mapping and network stability through residual connections and random depth mechanisms.

[0142] Finally, the features output by the deep GCN are mapped to the classification space through global average pooling and convolution operations, generating the final prediction results. The classification head is a part of the model that maps learned feature representations to class predictions. It is usually located at the end of the network, following the feature extraction layer. The features output by the deep GCN are mapped to the classification space through global average pooling and convolution operations in the classification head, generating the final prediction results. Exemplarily, the classification space is transformed into scores for each class through a fully connected layer or a 1x1 convolution operation, resulting in the prediction results.

[0143] Fault detection of transmission lines is essential for ensuring the reliable operation of power systems. The embodiments of the present application aim to address the problem that fine features of transmission lines in complex environments can cause local feature extraction to be affected by local noise. A deep learning-based transmission line detection technology is proposed. A subsampling module is proposed to speed up the calculation rate. An SPD-Conv module is proposed to enhance the model's feature extraction capability for low-resolution images and small targets. An NAM attention mechanism is proposed to efficiently improve the model's performance. Compared with the widely used models, this method has certain advantages in comprehensive performance while maintaining lightweight, and can well complete the transmission line fault detection task. The main contributions include the following points:

[0144] To address the problem of slow training cycle during model training, a subsampling framework is established to speed up the calculation efficiency while preserving the original detection accuracy.

[0145] To address the problem of image blurring and small detection targets in transmission lines, a space-to-depth convolution (SPD-Conv) is used to establish a detection framework that can effectively identify low-resolution images and small targets.

[0146] To address the problem of low transmission line detection accuracy, a normalized attention module (NAM) is introduced to enhance the detection effect while keeping the model size small and the computational complexity low, enabling accurate identification of transmission lines in complex environments.

[0147] The effect of the above method is verified by specific simulation experiments. The experiments are carried out under the pytorch framework, and the parameters of the software and hardware platforms are shown in Table 1.

[0148] Table 1

[0149]

[0150] (1) Construct the data set. The data set used in the experiment is composed of images obtained by shooting the power transmission line by a drone, including insulators, bird nests and other obstacles, etc. The bird nest and insulator training set and verification set are used in the embodiment of the application, and the data set has a total of 6295 pictures. The specific classification is shown in Table 2, and Table 2 is a schematic diagram of the number of data set annotations.

[0151] Table 2

[0152]

[0153] (2) Evaluation index. The following performance indicators are given for the performance evaluation of the power transmission line fault detection of the embodiment of the application, which are used to verify the effectiveness and accuracy of the model:

[0154] Precision represents the probability of correct prediction, and the calculation formula is as follows: . Where TP represents the number of true positives, and FP represents the number of false positives.

[0155] Recall represents the probability of correct prediction in the total amount of accurate detection and missed detection, and the calculation formula is as follows: ; Where FN represents the number of false negatives.

[0156] Mean Average Precision (MAP) is a measure of algorithm accuracy, which is the average of the accuracy rate. The calculation formula is as follows: . Where, is the precision value under the recall .

[0157] (3) Experimental results and analysis. In order to verify the advantages of SSA-Vision Gnn compared with Vision Gnn algorithm in power transmission line fault detection, the Vision Gnn algorithm and the SSA-Vision Gnn algorithm are compared.

[0158] Figure 10 is a schematic diagram of the relationship between the precision rate and the training period of the power transmission line fault detection under the Vision Gnn algorithm and the SSA-Vision Gnn algorithm,Figure 11 Figure 6 is a schematic diagram of the relationship between the recall rate and the training period for the Vision Gnn algorithm and the SSA-Vision Gnn algorithm for fault detection of a power transmission line. Figure 12 Figure 7 is a schematic diagram of the relationship between the average precision and the training period for the Vision Gnn algorithm and the SSA-Vision Gnn algorithm for fault detection of a power transmission line when the intersection over union threshold is 0.5.

[0159] In Figure 10 to Figure 12 the horizontal axis is the training period (epoch), which refers to the process of passing the entire training data set through the training algorithm once.

[0160] Figure 10 Figure 8 is a comparison chart of the precision of different algorithms under complex background. From Figure 10 It can be seen that the precision of the two algorithms generally shows an upward trend. Among them, the Vision Gnn algorithm has low precision in the initial stage, then rapidly rises and reaches a precision close to 0.9 at about 20 epochs, and then has small fluctuations but basically stabilizes at about 0.9. The precision of the SSA-Vision Gnn algorithm is low in the initial stage, then its precision rises relatively smoothly, and reaches a precision close to 0.95 at about 40 epochs, and then basically stabilizes at about 0.95. The SSA-Vision Gnn algorithm is significantly better than the Vision Gnn algorithm in terms of precision, especially in the early rising stage, the SSA-Vision Gnn algorithm can reach a higher precision more quickly, and the final stable precision value is also higher. The SSA-vision Gnn algorithm adopts a more optimized strategy and algorithm structure, which enables it to learn features more effectively during training, thereby quickly improving the precision in the early stage and achieving a higher and more stable precision in the later stage.

[0161] Figure 11 Figure 9 is a comparison chart of the recall rate of different algorithms under complex background. From Figure 11It can be seen that the recall rates of the two algorithms overall show an upward trend. Among them, the Vision Gnn algorithm has a lower recall rate in the initial stage, then rises rapidly, reaches a recall rate close to 0.9 at about 20 epochs, and then has a small fluctuation and stabilizes at about 0.9. The recall rate of the SSA-Vision Gnn algorithm is higher than that of the Vision Gnn algorithm in the initial stage, and then its recall rate rises relatively smoothly, reaching a recall rate close to 0.95 at about 40 epochs, and then basically stabilizing at about 0.95. The SSA-Vision Gnn algorithm is obviously superior to the Vision Gnn algorithm in recall rate, especially in the early rising stage, the SSA-Vision Gnn algorithm can reach a higher recall rate more quickly, and the final stable recall rate value is also higher. The SSA-Vision Gnn algorithm uses a more effective feature extraction and matching mechanism, so that it can more comprehensively identify the target in the training process, thereby quickly improving the recall rate in the early stage, and achieving a higher and more stable recall rate in the later stage.

[0162] Figure 12 The average precision (Mean Average Precision at IoU threshold 0.5, MAP@0.5) of different algorithms under complex background when the intersection over union threshold is 0.5 is compared. From Figure 12 It can be seen that the MAP@0.5 of the two algorithms overall shows an upward trend. Among them, the MAP@0.5 of the Vision Gnn algorithm is low in the initial stage, then rises rapidly, reaches a MAP@0.5 close to 0.9 at about 20 epochs, and then has a small fluctuation and stabilizes at about 0.9. The MAP@0.5 of the SSA-Vision Gnn algorithm is higher than that of the Vision Gnn algorithm in the initial stage, and then its MAP@0.5 rises relatively smoothly, reaching a MAP@0.5 close to 0.95 at about 40 epochs, and then basically stabilizing at about 0.95. The SSA-Vision Gnn algorithm is obviously superior to the Vision Gnn algorithm in MAP@0.5, especially in the early rising stage, the SSA-Vision Gnn algorithm can reach a higher MAP@0.5 more quickly, and the final stable MAP@0.5 value is also higher. The SSA-Vision Gnn algorithm has a better design in model structure and parameter optimization, so that it can more efficiently learn and identify the target in the training process, thereby quickly improving the MAP@0.5 in the early stage, and achieving a higher and more stable MAP@0.5 in the later stage.

[0163] In order to further verify the effectiveness of the improved algorithm proposed in the embodiment of the present application, the improved algorithm is compared with other commonly used models, as shown in Table 3. Table 3 shows the results of various indicators of different models when trained on the same data set.

[0164] Table 3

[0165]

[0166] Table 3 compares the experimental results of various models across various metrics. Table 3 shows that the different models differ in terms of Floating Point Operations in Giga (FLOPs(G)), accuracy, recall, and MAP@0.5. SSA-Vision Gnn has a significant advantage in FLOPs(G), using only 5.1G, far lower than the other models. In terms of accuracy, SSD achieves the highest accuracy of 96.52%, while SSA-Vision Gnn achieves 92.11%, which is not the highest but still impressive. In terms of recall, SSA-Vision Gnn performs best at 90.59%. In terms of MAP@0.5, SSA-Vision Gnn performs well, reaching 89.85%. Overall, SSA-Vision Gnn maintains high accuracy, recall, and MAP@0.5 while maintaining low computational complexity. SSA-Vision Gnn optimizes its model design and adopts a more efficient structure and algorithm, enabling it to achieve better detection performance with fewer computing resources, thereby maintaining high accuracy, recall and MAP@0.5 with lower FLOPs.

[0167] Experimental results show that the SSA-Vision Gnn model maintains high detection accuracy while effectively reducing computational overhead when processing power line image detection tasks. By introducing subsampling, the model's inference speed is significantly improved compared to the original model; the SPD-Conv module effectively enhances the detection capabilities of low-resolution images and small targets, significantly improving the average accuracy of target detection; and the NAM attention module further optimizes the feature extraction effect, allowing the model to maintain a lightweight design while significantly improving detection accuracy. Overall, the improved solution proposed in the embodiment of this application can achieve efficient and accurate power line target detection in actual application scenarios.

[0168] Based on the same inventive concept, the embodiments of the present application further provide a power transmission line fault detection device for implementing the power transmission line fault detection method described above. The implementation scheme for solving the problem provided by the device is similar to the implementation scheme described in the above method, so the specific limitations in one or more power transmission line fault detection device embodiments provided below can refer to the limitations of the power transmission line fault detection method described above, which will not be repeated here.

[0169] In an exemplary embodiment, the power transmission line fault detection device comprises: a space-to-depth convolutional network unit; an initial image obtained by photographing the power transmission line is obtained, and the initial image is processed by using the space-to-depth convolutional network to obtain a target image; a normalization unit, the target image is processed by using a normalization attention module to obtain an output image; a subsampling unit, the output image is sampled and processed by using a subsampling module to obtain a to-be-processed image; a classification unit, the to-be-processed image is mapped to a classification space to generate a fault detection result of the power transmission line.

[0170] The above device embodiment description is similar to the description of the above method embodiments, and has similar beneficial effects as the method embodiments. For technical details not disclosed in the device embodiments of the present application, please refer to the description of the method embodiments of the present application for understanding.

[0171] Each module in the above power transmission line fault detection device can be realized by software, hardware and their combinations in whole or in part. The above modules can be embedded in or independent of the processor in the computer device in hardware form, or can be stored in the memory in the computer device in the form of software, so as to call and execute the operations corresponding to the above modules by the processor.

[0172] In an exemplary embodiment, Figure 13A structural diagram of a computer device provided in some embodiments is shown in the figure, which includes a processor, a memory, an input / output interface, a communication interface, a display unit and an input device. The processor, the memory and the input / output interface are connected through a system bus, and the communication interface, the display unit and the input device are connected to the system bus through the input / output interface. The processor of the computer device is configured to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for running the operating system and the computer program in the non-volatile storage medium. The input / output interface of the computer device is configured to exchange information between the processor and external devices. The communication interface of the computer device is configured to perform wired or wireless communication with external terminals, and the wireless communication can be achieved through Wireless Fidelity (WIFI), mobile cellular network, Near Field Communication (NFC) or other technologies. The computer program is executed by the processor to implement a power transmission line fault detection method. The display unit of the computer device is configured to form a visually visible picture, which can be a display screen, a projection device or a virtual reality imaging device. The display screen can be a liquid crystal display screen or an electronic ink display screen, and the input device of the computer device can be a touch layer overlaid on the display screen, or a key, trackball or touchpad arranged on the shell of the computer device, or an external keyboard, touchpad or mouse, etc.

[0173] Those skilled in the art can understand that, Figure 10 The structure shown in the figure is only a block diagram of part of the structure related to the scheme of the present application, and does not constitute a limitation on the computer device to which the scheme of the present application is applied. The specific computer device can include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0174] For example, the computer device includes a memory and a processor, the memory stores a computer program, and the processor executes the computer program to implement the steps of the method of any of the above embodiments.

[0175] In one embodiment, a computer readable storage medium is provided, and a computer program is executed by the processor to implement the steps of the method provided in any of the above embodiments.

[0176] In one embodiment, a computer program product is provided, which includes a computer program, and the computer program is executed by the processor to implement the steps of the method provided in any of the above embodiments.

[0177] Those skilled in the art will understand that all or part of the processes in the above-mentioned embodiment methods can be implemented by instructing related hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods.

[0178] The processor, each functional module or each functional unit in any embodiment of the present application may include any one or more of the following integrations: a general-purpose processor, an application-specific integrated circuit (ASIC), a digital signal processor (DSP), a digital signal processing device (DSPD), a programmable logic device (PLD), a field programmable gate array (FPGA), a central processing unit (CPU), a graphics processing unit (GPU), an embedded neural network processing unit (NPU), a controller, a microcontroller, a microprocessor, a programmable logic device, a discrete gate or transistor logic device, a discrete hardware component, a data processing logic based on quantum computing, an artificial intelligence (AI) processor, etc. The general-purpose processor may be a microprocessor or the processor may also be any conventional processor, etc.

[0179] The memory or computer readable storage medium in any of the embodiments of the present application can include at least one of a non-volatile memory and a volatile memory. The non-volatile memory includes integration of one or more of the following: Read Only Memory (ROM), Programmable Read-Only Memory (PROM), Erasable Programmable Read-Only Memory (EPROM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Ferromagnetic Random Access Memory (FRAM), Flash Memory, magnetic random access memory, optical disk, Compact Disc Read-Only Memory (CD-ROM), magnetic tape, floppy disk, flash memory, optical storage, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetoresistive random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, volatile memory, and the like. The volatile memory includes integration of one or more of the following: Random Access Memory (RAM) or external cache memory, and the like. As an illustration but not limitation, the RAM can be in various forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM), and the like.

[0180] Any of the technical features of the above embodiments can be combined, and for the sake of brevity, not all possible combinations are described, however, any combination of the technical features is considered to be within the scope of the present application as long as there is no contradiction.

[0181] The above embodiments only express several implementation ways of the present application, and the description is specific and detailed, but it should not be understood as a limitation to the patent scope of the present application. It should be pointed out that for ordinary skilled in the art, without departing from the concept of the present application, several modifications and improvements can be made, which all belong to the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the appended claims.

Claims

1. A method for detecting a power transmission line fault, characterized in that: The method comprises: Acquire an initial image obtained by photographing the transmission line, and process the initial image using a space-to-depth convolutional network to obtain a target image; Processing the target image using a normalized attention module to obtain an output image; Using a subsampling module to perform sampling processing on the output image to obtain an image to be processed; The image to be processed is mapped to a classification space to generate a fault detection result of the power transmission line.

2. The method according to claim 1, characterized in that The process of processing the initial image using a space-to-depth convolutional network to obtain a target image includes: Dividing the initial image into a plurality of sub-images, and extracting pixel data having the same position from each of the plurality of sub-images; Merging the extracted pixel data to obtain a plurality of rearranged images; Splicing the multiple rearranged images in the channel dimension to obtain a dimensionally spliced ​​image; Perform convolution processing on the dimensional spliced ​​image to obtain the target image.

3. The method according to claim 1, characterized in that The process of processing the target image using the normalized attention module to obtain an output image includes: Normalizing the image of each channel in the target image to obtain normalized image data of each channel, and scaling and translating the normalized image data of each channel to obtain adjusted image data of each channel; Normalizing the initial weight of each channel to obtain the normalized weight of each channel, and multiplying the adjusted image data of each channel by the normalized weight of each channel to obtain the initial channel enhanced image of each channel; Normalizing the initial channel-enhanced image of each channel using an activation function to obtain a channel-enhanced feature image; Normalizing the weight of each pixel in the channel-enhanced feature image in the depth dimension to obtain a normalized weight for each depth, and multiplying the channel-enhanced feature image and the normalized weight for each depth to obtain pixel data for each depth; An activation function is used to normalize the pixel data at each depth to obtain the output image.

4. The method according to any one of claims 1 to 3, characterized in that The adopting a subsampling module to perform sampling processing on the output image to obtain an image to be processed includes: dividing the output image into a plurality of image blocks, and generating an initial graph structure based on the plurality of image blocks; Performing feature aggregation and feature updating on the initial graph structure to obtain a graph-level processed graph structure, and fusing the initial graph structure and the graph-level processed graph structure to obtain a fused graph structure; Processing the fused graph structure using a feedforward neural network to obtain a transformed graph structure, and fusing the fused graph structure and the transformed graph structure to obtain a target graph structure; The target graph structure is sampled to obtain a sampling graph structure, and the image to be processed is determined according to the sampling graph structure.

5. The method according to claim 4, characterized in that Generating an initial graph structure according to the plurality of image blocks includes: Performing feature extraction on each of the image blocks to obtain a feature vector of each of the image blocks, and mapping the feature vector of each of the image blocks to each node in a graph structure to obtain a node set; Calculating the similarity between each node and other nodes, and determining the nodes corresponding to the largest preset number of similarities as adjacent nodes of each node; Determine the directed edges from the neighboring nodes of each of the nodes to each of the nodes as an edge set; The node set and the edge set are combined to generate the initial graph structure.

6. The method according to claim 4, characterized in that The performing feature aggregation and feature updating on the initial graph structure to obtain a graph-level processed graph structure includes: Determining relative eigenvectors between the eigenvector of each node in the initial graph structure and the eigenvectors of each adjacent node of the node; Extracting the maximum value in each dimension from each relative feature vector to obtain the maximized feature of each node; Performing feature aggregation on the feature vector of each node and the maximized aggregation feature of each node to obtain an aggregated feature of each node; The aggregated features of each of the nodes are updated using the updated weights to obtain the graph-level processing graph structure.

7. The method according to claim 4, characterized in that The sampling process of the target graph structure to obtain a sampled graph structure includes: Obtaining the selectivity of each node in the target graph structure, and determining the nodes corresponding to the selectivity greater than a threshold as key nodes; According to the target graph structure, the key node and all adjacent nodes of the key node are determined as sampling nodes; The sampling nodes and the directed edges between the sampling nodes are combined to form the sampling graph structure.

8. The method according to any one of claims 1 to 3, characterized in that Mapping the image to be processed to a classification space to generate a fault detection result of the power transmission line includes: Performing global average pooling processing on the image to be processed to obtain a global average pooling feature vector; Performing convolution processing on the global average pooling feature vector to obtain a convolution feature vector; The convolution feature vector is mapped to a classification space to generate a fault detection result of the transmission line.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 8 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 8 are implemented.

Images