Power transmission line icing detection method and system

Abstract

The invention provides a power transmission line icing detection method and system. The method comprises the following steps: collecting an icing image, temperature distribution data and sparse point cloud data; performing alignment processing on the sparse point cloud data to generate a distance mapping graph; inputting the icing image into a SegFormer model, and adjusting the expansion rate of a grouping depth separable convolutional layer through a distance mapping graph to obtain a first icing segmentation feature; generating an icing area pseudo label based on the temperature distribution data, and supervising the icing image to be segmented into a second icing segmentation feature through comparative learning; fusing and outputting the first icing segmentation feature and the second icing segmentation feature to obtain an icing feature; and according to the icing characteristics, extracting and complementing the corresponding sparse point cloud data, and constructing a three-dimensional model of power transmission line icing to obtain a power transmission line icing detection result, so that the technical problem of low accuracy of video detection of the icing condition of the power transmission line under the condition of severe weather in the prior art can be solved.

Description

Technical Field

[0001] The present invention relates to the technical field of transmission line detection, and in particular to a method and system for detecting icing on a transmission line. Background Art

[0002] Icing is a widespread natural phenomenon. For transmission lines, severe icing may cause failures and even lead to catastrophic accidents such as large-scale power outages.

[0003] Currently, workers can use real-time video monitoring to monitor transmission line conditions, allowing them to take timely de-icing measures when ice thickens. However, this monitoring method has the drawback of requiring a large number of workers to remain in front of the monitors for extended periods, resulting in a waste of human resources. Furthermore, in inclement weather, ice and rain can blur the video camera's image, affecting the accuracy of determining line ice coverage. Summary of the Invention

[0004] Based on this, the purpose of the present invention is to provide a method and system for detecting icing on transmission lines, so as to solve the problem in the background art of low accuracy in video detection of icing on transmission lines in severe weather conditions.

[0005] In one aspect, the present invention provides a method for detecting ice coating on a transmission line, the method comprising: Collect ice coverage images, temperature distribution data, and sparse point cloud data of transmission lines; performing alignment processing on the sparse point cloud data to generate a distance map aligned with the ice cover image at the pixel level; Inputting the ice-covered image into a SegFormer model for feature extraction, and adjusting the expansion rate of the grouped depth-separable convolutional layer in the SegFormer model through a distance map to obtain a first ice-covered segmentation feature; Generate a pseudo label of the ice-covered area based on the temperature distribution data, and segment the ice-covered image into a second ice-covered segmentation feature through comparative learning supervision; The first ice segmentation feature and the second ice segmentation feature are fused and outputted through a gated attention mechanism to obtain an ice feature; According to the icing features, the corresponding sparse point cloud data is extracted and completed, a three-dimensional model of the transmission line icing is constructed, and the transmission line icing detection result is obtained.

[0006] Compared with the prior art, the beneficial effects of the present invention are: generating a distance map aligned with the pixel level of the ice-covered image through sparse point cloud data, inputting the ice-covered image into the SegFormer model for feature extraction, and adjusting the expansion rate of the grouped depth-separable convolutional layer in the SegFormer model through the distance map to obtain the first ice-covered segmentation feature, thereby improving the precision and accuracy of edge detection, generating a pseudo-label of the ice-covered area based on the temperature distribution data, and supervising the segmentation of the ice-covered image into the second ice-covered segmentation feature through comparative learning, analyzing the ice-covered situation more comprehensively from a thermal perspective, reflecting the ice-covered state, and complementing the first ice-covered cutting feature to jointly improve the extraction and analysis of ice-covered features, which can effectively reduce the impact of severe weather and improve the precision and accuracy of detection, thereby solving the technical problem of low accuracy in video detection of ice-covered transmission lines in severe weather.

[0007] According to one aspect of the above technical solution, the SegFormer model includes an encoder part and a decoder part, which are fused through an MLP network. The encoder part includes four layered MiTBlocks, and the MiT Block includes an MSA layer, an improved Mix-FFN layer, and overlapping patch merging. The improved Mix-FFN layer includes an improved layer, a GELU layer, a depth convolution layer, an FC layer, and a residual connection layer connected in sequence, wherein the improved layer replaces the FC layer with a Ghost convolution layer.

[0008] According to one aspect of the above technical solution, the step of adjusting the expansion rate of the grouped depth-separable convolutional layer in the SegFormer model through the distance map specifically includes: The distance map is mapped to the expansion rate of the grouped depth-separable convolution layer in the Ghost convolution layer through a learnable parameter mapping function. The formula is as follows: , in, is the expansion rate, is the preset maximum expansion rate, is the sigmoid function, is the RELU activation function, 、 is the weight matrix, 、 is the bias term, is the distance value in the distance map; The symmetrical filling amount is calculated based on the expansion ratio as follows: , in, is the size of the convolution kernel in the group depth-wise separable convolution layer, To round down, Symmetrical filling amount.

[0009] According to one aspect of the above technical solution, the step of generating a pseudo label of the ice-covered area based on the temperature distribution data and segmenting the ice-covered image into a second ice-covered segmentation feature through comparative learning supervision specifically includes: Based on the temperature distribution data, the ice-covered image is divided into different regions, and pseudo labels of ice-covered sub-regions are generated for different regions. The temperature gradient and temperature variance between ice-covered sub-regions and non-ice-covered sub-regions in the same region are calculated, and the probability of pseudo labels of ice-covered sub-regions is calculated. The formula is as follows: , in, The first The probability of a pseudo label, 、 is a hyperparameter, The first The temperature gradient between the point and the non-iced sub-area, is the maximum temperature gradient between the iced sub-region and the non-iced sub-region, The first Temperature variance between the point and the non-iced sub-area; According to the probability of the pseudo-label of the icing sub-region, the icing sub-regions in the same area are divided into strong icing sub-regions and weak icing sub-regions, and feature constraints are performed to optimize the strong icing sub-regions and weak icing sub-regions, respectively, to construct positive and negative sample pairs for contrastive learning. The formula is as follows: , in, 、 are two adjacent pixels in the strong icing sub-region 、 Features, is the temperature characteristic coefficient of the strong icing sub-region, 、 Two adjacent pixels 、 temperature, 、 are two adjacent pixels in the weak icing sub-region. 、 Features, is the temperature characteristic coefficient of the weakly icing sub-region; Obtain ice sub-regions with the same probability in different regions, calculate the difference between strong ice sub-regions and weak ice sub-regions across regions, and segment them into second ice segmentation features based on positive and negative sample pairs of supervised contrast learning based on the difference.

[0010] According to one aspect of the above technical solution, the difference includes the difference between two heavily iced sub-regions across regions, the difference between two weakly iced sub-regions across regions, and the difference between a heavily iced region and a weakly iced region across regions.

[0011] According to one aspect of the above technical solution, the step of fusing the first ice segmentation feature and the second ice segmentation feature through a gated attention mechanism to obtain an ice feature specifically includes: The first ice segmentation feature and the second ice segmentation feature are channel-joined, and a spatial attention weight map is generated through the Sigmoid function. The formula is as follows: , in, is the spatial attention weight map, is the sigmoid function, 、 They are the first ice-covered cutting feature and the second ice-covered cutting feature, Indicates channel splicing processing, Represents convolution compression processing; Based on the spatial attention weight map, the first ice-covered cutting feature and the second ice-covered cutting feature are fused to output the ice-covered feature. The formula is as follows: , in, The ice cover characteristic is Represents pixel-by-pixel multiplication.

[0012] According to one aspect of the above technical solution, the steps of extracting and completing the corresponding sparse point cloud data based on the icing features, constructing a three-dimensional model of the transmission line icing, and obtaining the transmission line icing detection results specifically include: Extracting the corresponding sparse point cloud data based on the ice cover features, constructing a three-dimensional model based on Euclidean distance and normal vector similarity, and completing the three-dimensional model based on a graph neural network; Based on the three-dimensional model, the distribution of ice thickness on the transmission line is calculated, and a risk warning signal is output according to the distribution of ice thickness on the transmission line.

[0013] According to one aspect of the above technical solution, the similarity calculation formula is: , in, for Point and The similarity of the points, is the balance coefficient, 、 They are point, The coordinates of the point, 、 They are point, The normal vector of the point.

[0014] According to one aspect of the above technical solution, the step of calculating the ice thickness distribution of the transmission line based on the three-dimensional model specifically includes: Fit a cylindrical model to the surface of the un-iced transmission line to obtain the radius and reference plane coordinates of the transmission line; Based on the radius of the transmission line and the reference plane coordinates, the ice thickness distribution of the transmission line is calculated using the following formula: , in, is the radius, 、 is the reference plane coordinate, 、 For the The coordinates of the point, For the Ice thickness at a point.

[0015] Another aspect of the present invention provides a transmission line icing detection system, the system being configured to execute the above-mentioned transmission line icing detection method, the system comprising: Data acquisition module, used to collect ice coverage images, temperature distribution data, and sparse point cloud data of transmission lines; a data processing module, configured to perform alignment processing on the sparse point cloud data to generate a distance map aligned with the ice cover image at the pixel level; A first feature extraction module is configured to input the ice-covered image into a SegFormer model for feature extraction, and adjust the expansion rate of the grouped depth-separable convolutional layer in the SegFormer model through a distance map to obtain a first ice-covered segmentation feature; A second feature extraction module is used to generate a pseudo label of the ice-covered area based on the temperature distribution data, and to supervise the segmentation of the ice-covered image into a second ice-covered segmentation feature through comparative learning; a feature fusion module, configured to fuse and output the first ice segmentation feature and the second ice segmentation feature through a gated attention mechanism to obtain an ice feature; The result output module is used to extract and complete the corresponding sparse point cloud data according to the ice coverage characteristics, construct a three-dimensional model of the transmission line ice coverage, and obtain the transmission line ice coverage detection result. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 This is a flow chart of a method for detecting ice coating on a transmission line in a first embodiment of the present invention; Figure 2 This is a structural block diagram of a power transmission line ice detection system in a second embodiment of the present invention; Description of the accompanying symbols: Data acquisition module 100, data processing module 200, first feature extraction module 300, second feature extraction module 400, feature fusion module 500, result output module 600; The following specific embodiments will further illustrate the present invention in conjunction with the above-mentioned drawings. DETAILED DESCRIPTION

[0017] To facilitate understanding of the present invention, the present invention will be described more fully below with reference to the accompanying drawings. The drawings illustrate several embodiments of the present invention. However, the present invention may be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided to provide a more thorough and comprehensive understanding of the present invention.

[0018] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one skilled in the art to which this invention pertains. The terms used in this specification of the present invention are for the purpose of describing specific embodiments only and are not intended to limit the present invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0019] Example 1 See also Figure 1 , which shows a method for detecting ice coating on a transmission line in a first embodiment of the present invention, including steps S10 to S15.

[0020] Step S10, collecting ice coverage images, temperature distribution data, and sparse point cloud data of the transmission line; Ice coverage images, temperature distribution data, and sparse point cloud data can be acquired through cameras, infrared sensors, and lidar, respectively. By acquiring and processing these multiple data sets, environmental interference (such as rain, fog, and nighttime conditions) can be effectively reduced, improving the accuracy and precision of transmission line inspections.

[0021] Step S11, performing alignment processing on the sparse point cloud data to generate a distance map aligned with the ice cover image at the pixel level; As an example but not a limitation, the distance map can be projected onto the image coordinate system through sparse point cloud data, and bilinear interpolation is used to generate a distance map consistent with the resolution of the ice-covered image. For the position or angle offset between the sparse point cloud data and the ice-covered image, an online iterative nearest point algorithm is used for real-time correction to ensure pixel-level alignment of the distance map and the image, thereby improving the accuracy of transmission line detection.

[0022] Step S12: inputting the ice-covered image into a SegFormer model for feature extraction, and adjusting the expansion rate of the grouped depth-separable convolutional layer in the SegFormer model through a distance map to obtain a first ice-covered segmentation feature; The SegFormer model includes an encoder part and a decoder part, which are fused through an MLP network. The encoder part includes four layered MiT Blocks, which include an MSA layer, an improved Mix-FFN layer, and overlapping patch merging. Furthermore, the SegFormer model is a segmentation transformer, which is the abbreviation of "Segmentation Transformer". It is a model architecture for semantic segmentation tasks. MLP network (Multi-Layer Perceptron): A multi-layer perceptron network is a feedforward neural network composed of multiple neuron layers that can perform nonlinear transformation and feature learning on input data. MiT Block (Mix Transformer Block): A mixed transformer module, which is the core component of the encoder in the SegFormer model and is used for feature extraction and processing. MSA layer (Multi-Head Self-Attention): A multi-head self-attention layer, an attention mechanism that calculates attention in parallel through multiple heads, can more comprehensively capture the relationship between input features. Mix-FFN layer (Mixed Feed-Forward Network): A mixed feedforward network layer that combines multiple operations to transform and process features to improve the feature expression ability of the model.

[0023] Furthermore, the four MiT Blocks in the encoder extract features at different scales from the input image, with resolutions of 1 / 4, 1 / 8, 1 / 16, and 1 / 32 of the original ice image. In the decoder, an MLP network extracts high-resolution, low-level semantic information from the decoder structure and fuses it with the low-resolution, high-level semantic information. This ensures that the SegFormer model achieves more accurate ice edge segmentation and optimizes ice segmentation results.

[0024] The improved Mix-FFN layer includes an improved layer, a GELU layer, a depth convolution layer, an FC layer, and a residual connection layer connected in sequence, and the improved layer replaces the FC layer with a Ghost convolution layer.

[0025] Furthermore, the FC layer (Fully Connected Layer) is a fully connected layer. The Ghost convolution layer replaces the FC layer and can generate redundant features through a small amount of conventional convolution and linear operations, reduce the number of parameters, achieve lightweight, and enhance feature diversity through the Ghost convolution layer.

[0026] Specifically, the Ghost convolutional layer uses a small number of regular convolutions to generate basic features. It then generates redundant features through inexpensive operations (such as depthwise or pointwise convolutions), concatenating the basic and redundant features while maintaining the same number of output channels. The Ghost module generates features in steps, avoiding the high computational cost of fully connected layers.

[0027] At the same time, the depthwise convolutional layer and the FC layer after the depthwise convolutional layer remain unchanged, ensuring that the position information encoding capability of the improved Mix-FFN layer is not lost.

[0028] In addition, the input and output dimensions (number of channels, resolution) are exactly the same as the original Mix-FFN layer, which can directly replace the original module without adjusting the model structure, and the structure is simple.

[0029] In this embodiment, the Ghost convolution layer includes a main path and an auxiliary path. The main path uses 1×1 regular convolution to generate basic features, and the auxiliary path uses grouped depth convolution and 1×1 point-by-point convolution to splice the outputs of the main path and the auxiliary path along the channel dimension.

[0030] Among them, the grouped depth in the Ghost convolution layer can be separated into convolution layers. The convolution kernel expansion rate of the Ghost convolution layer is dynamically adjusted in combination with the distance map to expand the receptive field of the edge area, improve the extraction ability of the edge area, and enhance the segmentation accuracy.

[0031] Specifically, the distance map is mapped to the expansion rate of the grouped depth-separable convolutional layer through a learnable parameter mapping function. The formula is as follows: , in, is the expansion rate, is the preset maximum expansion rate, is the sigmoid function, is the RELU activation function, 、 is the weight matrix, 、 is the bias term, is the distance value in the distance map; The symmetrical filling amount is calculated based on the expansion ratio as follows: , in, is the size of the convolution kernel in the group depth-wise separable convolution layer, To round down, Symmetrical filling amount.

[0032] Step S13, generating a pseudo label of the ice-covered area based on the temperature distribution data, and segmenting the ice-covered image into a second ice-covered segmentation feature through comparative learning supervision; Specifically, the ice-covered image is divided into different regions based on the temperature distribution data, and pseudo labels of ice-covered sub-regions are generated for different regions. The temperature gradient and temperature variance between ice-covered sub-regions and non-ice-covered sub-regions in the same region are calculated, and the probability of pseudo labels of ice-covered sub-regions is calculated. The formula is as follows: , in, The first The probability of a pseudo label, 、 is a hyperparameter, The first The temperature gradient between the point and the non-iced sub-area, is the maximum temperature gradient between the iced sub-region and the non-iced sub-region, The first Temperature variance between the point and the non-iced sub-area; According to the probability of the pseudo-label of the iced sub-region, the iced sub-regions in the same area are divided into strong iced sub-regions and weak iced sub-regions, and feature constraints are performed to optimize the strong iced sub-regions and weak iced sub-regions. The formula is as follows: , in, 、 are two adjacent pixels in the strong icing sub-region 、 Features, is the temperature characteristic coefficient of the strong icing sub-region, 、 Two adjacent pixels 、 temperature, 、 are two adjacent pixels in the weak icing sub-region. 、 Features, is the temperature characteristic coefficient of the weakly icing sub-region; In order to avoid relying solely on data-driven contrastive learning, which may ignore the physical characteristics of ice cover (such as sudden changes in edge temperature), feature constraints are added to enable the model to learn feature representations that conform to the laws of thermodynamics, enhance the physical rationality of edge segmentation, and further improve detection accuracy.

[0033] Obtaining ice-covered sub-regions with the same probability in different regions, calculating the difference between the strong ice-covered sub-regions and the weak ice-covered sub-regions across the regions, and segmenting the positive and negative sample pairs of the supervised contrastive learning based on the difference into a second ice-covered segmentation feature; The difference includes the difference between two heavily iced sub-regions across a region, the difference between two weakly iced sub-regions across a region, and the difference between a heavily iced region and a weakly iced region across a region.

[0034] By dividing the area and the pseudo-label probability of the ice-covered sub-area, the difference is calculated to reduce the impact of the supervisory signal in the unreliable area and improve the model's anti-interference ability.

[0035] It should be noted that the contrastive learning of positive sample pairs in the same region: pixel pairs within the strong ice sub-region or the weak ice sub-region, enhances intra-class consistency and captures local gradient features; Comparative learning of negative sample pairs in the same region: pixel pairs between strong and weak ice-covered sub-regions to enhance inter-class discrimination; Contrastive learning of positive sample pairs across regions: two strongly iced sub-regions across regions, two weakly iced sub-regions across regions; Contrastive learning of negative sample pairs across regions: strong ice-covered areas and weak ice-covered areas across regions force similar types across regions to move closer and different types to move away.

[0036] Furthermore, a joint cross-region and intra-region comparative learning framework is proposed, balancing local details with global generalization. Differentiated comparative supervision is designed to enhance the model's adaptability to complex scenarios and improve the accuracy and robustness of ice cover recognition.

[0037] Step S14: fusing the first ice segmentation feature and the second ice segmentation feature through a gated attention mechanism to obtain an ice feature; Specifically, the first ice segmentation feature and the second ice segmentation feature are channel-joined, and a spatial attention weight map is generated through the Sigmoid function. The formula is as follows: , in, is the spatial attention weight map, is the sigmoid function, 、 They are the first ice-covered cutting feature and the second ice-covered cutting feature, Indicates channel splicing processing, Represents convolution compression processing; Based on the spatial attention weight map, the first ice-covered cutting feature and the second ice-covered cutting feature are fused to output the ice-covered feature. The formula is as follows: , in, The ice-covered feature Represents pixel-by-pixel multiplication.

[0038] Among them, the first ice-covered cutting feature improves the precision and accuracy of edge detection through sparse point cloud data. The second ice-covered cutting feature uses temperature distribution data to more comprehensively analyze the ice situation from a thermal perspective, which can reflect the ice status (such as thickness, degree of melting, etc.). It complements the first ice-covered cutting feature and jointly improves the extraction and analysis of ice features.

[0039] Step S15: extract and complete the corresponding sparse point cloud data according to the icing features, construct a three-dimensional model of the transmission line icing, and obtain the transmission line icing detection result.

[0040] Specifically, according to the ice cover features, the corresponding sparse point cloud data is extracted, and a three-dimensional model is constructed based on the Euclidean distance and the similarity of the normal vector, and the three-dimensional model is completed based on the graph neural network; The formula for calculating similarity is: , in, for Point and The similarity of the points, is the balance coefficient, 、 They are point, The coordinates of the point, 、 They are point, The normal vector of the point.

[0041] Based on the three-dimensional model, the distribution of ice thickness on the transmission line is calculated, and a risk warning signal is output according to the distribution of ice thickness on the transmission line.

[0042] Specifically, a cylindrical model is fitted to the surface of the un-ice-covered transmission line to obtain the radius and reference plane coordinates of the transmission line; Based on the radius of the transmission line and the reference plane coordinates, the ice thickness distribution of the transmission line is calculated using the following formula: , in, is the radius, 、 is the reference plane coordinate, 、 For the The coordinates of the point, For the Ice thickness at a point.

[0043] In summary, the transmission line icing detection method in the above-mentioned embodiment of the present invention generates a distance map aligned with the pixel level of the icing image through sparse point cloud data, inputs the icing image into the SegFormer model for feature extraction, and adjusts the expansion rate of the grouped depth-separable convolution layer in the SegFormer model through the distance map to obtain the first icing segmentation feature, thereby improving the precision and accuracy of edge detection, and generating pseudo labels of icing areas based on temperature distribution data. The icing image is segmented into a second icing segmentation feature through comparative learning supervision, which analyzes the icing situation more comprehensively from a thermal perspective and reflects the icing state. It complements the first icing cutting feature to jointly improve the extraction and analysis of icing features and improve the precision and accuracy of detection, thereby solving the technical problem of low accuracy in video detection of icing conditions on transmission lines in severe weather conditions.

[0044] Example 2 See also Figure 2 , which shows a transmission line icing detection system in a second embodiment of the present invention, the system includes: The data acquisition module 100 is used to collect ice coverage images, temperature distribution data, and sparse point cloud data of the transmission line; A data processing module 200 is configured to perform alignment processing on the sparse point cloud data to generate a distance map aligned with the ice cover image at the pixel level; A first feature extraction module 300 is configured to input the ice-covered image into a SegFormer model for feature extraction, and adjust the dilation rate of the grouped depthwise separable convolutional layer in the SegFormer model using a distance map to obtain a first ice-covered segmentation feature; A second feature extraction module 400 is configured to generate a pseudo label for an ice-covered area based on the temperature distribution data, and to segment the ice-covered image into a second ice-covered segmentation feature through comparative learning supervision; A feature fusion module 500 is configured to fuse the first ice segmentation feature and the second ice segmentation feature through a gated attention mechanism to output an ice feature; The result output module 600 is used to extract and complete the corresponding sparse point cloud data according to the ice coverage characteristics, construct a three-dimensional model of the transmission line ice coverage, and obtain the transmission line ice coverage detection result.

[0045] In summary, the transmission line icing detection system in the above-mentioned embodiment of the present invention generates a distance map aligned with the pixel level of the icing image through sparse point cloud data, inputs the icing image into the SegFormer model for feature extraction, and adjusts the expansion rate of the grouped depth-separable convolution layer in the SegFormer model through the distance map to obtain the first icing segmentation feature, thereby improving the precision and accuracy of edge detection, and generating pseudo labels of icing areas based on temperature distribution data. The icing image is segmented into a second icing segmentation feature through comparative learning supervision, which analyzes the icing situation more comprehensively from a thermal perspective and reflects the icing status. It complements the first icing cutting feature to jointly improve the extraction and analysis of icing features and improve the precision and accuracy of detection, thereby solving the technical problem of low accuracy in video detection of icing conditions on transmission lines in severe weather conditions.

[0046] Throughout this specification, reference to terms such as "one embodiment," "some embodiments," "examples," "specific examples," or "some examples" means that a specific feature, structure, material, or characteristic described in conjunction with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, schematic representations of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in any one or more embodiments or examples.

[0047] The above-described embodiments merely illustrate several implementations of the present invention, and while their descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that a person skilled in the art would be able to make numerous variations and improvements without departing from the spirit of the present invention, all of which fall within the scope of protection of the present invention. Therefore, the scope of protection of the present invention shall be determined by the appended claims.

Claims

1. A method for detecting ice coating on a power transmission line, characterized in that: The method comprises: Collect ice coverage images, temperature distribution data, and sparse point cloud data of transmission lines; performing alignment processing on the sparse point cloud data to generate a distance map aligned with the ice cover image at the pixel level; Inputting the ice-covered image into a SegFormer model for feature extraction, and adjusting the expansion rate of the grouped depth-separable convolutional layer in the SegFormer model through a distance map to obtain a first ice-covered segmentation feature; Generate a pseudo label of the ice-covered area based on the temperature distribution data, and segment the ice-covered image into a second ice-covered segmentation feature through comparative learning supervision; The first ice segmentation feature and the second ice segmentation feature are fused and outputted through a gated attention mechanism to obtain an ice feature; According to the icing features, the corresponding sparse point cloud data is extracted and completed, a three-dimensional model of the transmission line icing is constructed, and the transmission line icing detection result is obtained.

2. The method for detecting ice coating on a power transmission line according to claim 1, wherein: The SegFormer model includes an encoder part and a decoder part, which are fused through an MLP network. The encoder part includes four layered MiT Blocks, which include an MSA layer, an improved Mix-FFN layer, and overlapping patch merging. The improved Mix-FFN layer includes an improved layer, a GELU layer, a depth convolution layer, an FC layer, and a residual connection layer connected in sequence, wherein the improved layer replaces the FC layer with a Ghost convolution layer.

3. The method for detecting ice coating on a power transmission line according to claim 2, wherein: The steps of adjusting the dilation rate of the grouped depthwise separable convolutional layer in the SegFormer model through the distance map include: The distance map is mapped to the expansion rate of the grouped depth-separable convolution layer in the Ghost convolution layer through a learnable parameter mapping function. The formula is as follows: ， in, is the expansion rate, is the preset maximum expansion rate, is the sigmoid function, is the RELU activation function, 、 is the weight matrix, 、 is the bias term, is the distance value in the distance map; The symmetrical filling amount is calculated based on the expansion ratio as follows: , in, is the size of the convolution kernel in the group depth-wise separable convolution layer, To round down, Symmetrical filling amount.

4. The method for detecting ice coating on a power transmission line according to claim 1, wherein: The steps of generating a pseudo label of the ice-covered area based on the temperature distribution data and segmenting the ice-covered image into a second ice-covered segmentation feature through comparative learning supervision specifically include: Based on the temperature distribution data, the ice-covered image is divided into different regions, and pseudo labels of ice-covered sub-regions are generated for different regions. The temperature gradient and temperature variance between ice-covered sub-regions and non-ice-covered sub-regions in the same region are calculated, and the probability of pseudo labels of ice-covered sub-regions is calculated. The formula is as follows: ， in, The first The probability of a pseudo label, 、 is a hyperparameter, The first The temperature gradient between the point and the non-iced sub-area, is the maximum temperature gradient between the iced sub-region and the non-iced sub-region, The first Temperature variance between the point and the non-iced sub-area; According to the probability of the pseudo-label of the icing sub-region, the icing sub-regions in the same area are divided into strong icing sub-regions and weak icing sub-regions, and feature constraints are performed to optimize the strong icing sub-regions and weak icing sub-regions, respectively, to construct positive and negative sample pairs for contrastive learning. The formula is as follows: ， in, 、 are two adjacent pixels in the strong icing sub-region 、 Features, is the temperature characteristic coefficient of the strong icing sub-region, 、 Two adjacent pixels 、 temperature, 、 are two adjacent pixels in the weak icing sub-region. 、 Features, is the temperature characteristic coefficient of the weakly icing sub-region; Obtain ice sub-regions with the same probability in different regions, calculate the difference between strong ice sub-regions and weak ice sub-regions across regions, and segment them into second ice segmentation features based on positive and negative sample pairs of supervised contrast learning based on the difference.

5. The method for detecting ice coating on a power transmission line according to claim 4, characterized in that: The difference includes the difference between two heavily iced sub-regions across a region, the difference between two weakly iced sub-regions across a region, and the difference between a heavily iced region and a weakly iced region across a region.

6. The method for detecting ice coating on a power transmission line according to claim 1, characterized in that: The step of fusing the first ice segmentation feature and the second ice segmentation feature through a gated attention mechanism to obtain an ice feature specifically includes: The first ice segmentation feature and the second ice segmentation feature are channel-joined, and a spatial attention weight map is generated through the Sigmoid function. The formula is as follows: ， in, is the spatial attention weight map, is the sigmoid function, 、 They are the first ice-covered cutting feature and the second ice-covered cutting feature, Indicates channel splicing processing, Represents convolution compression processing; Based on the spatial attention weight map, the first ice-covered cutting feature and the second ice-covered cutting feature are fused to output the ice-covered feature. The formula is as follows: ， in, The ice-covered feature Represents pixel-by-pixel multiplication.

7. The method for detecting ice coating on a power transmission line according to claim 1, wherein: The steps of extracting and completing corresponding sparse point cloud data according to the icing features, constructing a three-dimensional model of icing on the transmission line, and obtaining icing detection results on the transmission line specifically include: Extracting the corresponding sparse point cloud data based on the ice cover features, constructing a three-dimensional model based on Euclidean distance and normal vector similarity, and completing the three-dimensional model based on a graph neural network; Based on the three-dimensional model, the distribution of ice thickness on the transmission line is calculated, and a risk warning signal is output according to the distribution of ice thickness on the transmission line.

8. The method for detecting ice coating on a power transmission line according to claim 7, wherein: The formula for calculating similarity is: ， in, for Point and The similarity of the points, is the balance coefficient, 、 They are point, The coordinates of the point, 、 They are point, The normal vector of the point.

9. The method for detecting ice coating on a power transmission line according to claim 7, characterized in that: The step of calculating the ice thickness distribution of the transmission line based on the three-dimensional model specifically includes: Fit a cylindrical model to the surface of the un-iced transmission line to obtain the radius and reference plane coordinates of the transmission line; Based on the radius of the transmission line and the reference plane coordinates, the ice thickness distribution of the transmission line is calculated using the following formula: ， in, is the radius, 、 is the reference plane coordinate, 、 For the The coordinates of the point, For the Ice thickness at a point.

10. A transmission line ice detection system, characterized in that: The system is used to perform the transmission line icing detection method according to claims 1 to 9, and the system includes: Data acquisition module, used to collect ice coverage images, temperature distribution data, and sparse point cloud data of transmission lines; a data processing module, configured to perform alignment processing on the sparse point cloud data to generate a distance map aligned with the ice cover image at the pixel level; A first feature extraction module is configured to input the ice-covered image into a SegFormer model for feature extraction, and adjust the expansion rate of the grouped depth-separable convolutional layer in the SegFormer model through a distance map to obtain a first ice-covered segmentation feature; A second feature extraction module is used to generate a pseudo label of the ice-covered area based on the temperature distribution data, and to supervise the segmentation of the ice-covered image into a second ice-covered segmentation feature through comparative learning; a feature fusion module, configured to fuse and output the first ice segmentation feature and the second ice segmentation feature through a gated attention mechanism to obtain an ice feature; The result output module is used to extract and complete the corresponding sparse point cloud data according to the ice coverage characteristics, construct a three-dimensional model of the transmission line ice coverage, and obtain the transmission line ice coverage detection result.

Images