Method, system, device and medium for identifying icing thickness of power transmission line insulator

Abstract

The application provides a power transmission line insulator ice thickness identification method, system, device and medium, belongs to the technical field of computer vision and power facility state monitoring, and comprises the following steps: an improved high-resolution segmentation network is used as an encoder, an UperHead structure is used as a decoder, an ice-covered insulator image semantic segmentation model is constructed, a binary insulator segmentation image is obtained through model processing, and high-precision segmentation of an ice-covered insulator target under a complex interference background is realized. On the basis of high-precision target segmentation mask, physical quantitative area calculation is carried out, and the method has very high theoretical confidence and anti-noise effect. Based on end-to-end identification from a single image to final equivalent ice thickness measurement, the method has high accuracy and good robustness, and is suitable for various harsh anti-icing disaster inspection scenes. End-to-end quantitative evaluation from unmanned aerial vehicle inspection image to final ice thickness change detection result is realized, the method is efficient in calculation after deployment, and the result is intuitive and clear.

Description

Technical Field

[0001] This invention belongs to the field of computer vision and power facility condition monitoring technology, specifically relating to a method, system, equipment, and medium for identifying the ice thickness of insulators on transmission lines. Background Technology

[0002] Because transmission lines typically span long distances, traverse complex terrains, and are exposed to the elements for extended periods, they are highly susceptible to extreme natural geographical and microclimate conditions. In particular, high-altitude lines, while bearing their own mechanical stress, are further burdened by icing loads from low-temperature rain, snow, and freezing weather, significantly accelerating fatigue aging and mechanical damage to line components, posing serious hidden dangers to the power grid. Insulators, as numerous and crucial components in the lines, serving both mechanical support and electrical insulation functions, are highly prone to surface icing in low-temperature, high-humidity environments. The continuous accumulation of ice not only significantly alters the physical contours of the insulators and obscures their original texture but also drastically reduces their electrical performance, easily inducing severe faults such as "ice flashover" and "pollution flashover." In severe cases, this can lead to large-scale power outages or even regional power grid paralysis, affecting the safe and stable power supply for residential users, causing huge economic losses, and threatening the safe and stable operation of the power grid. Therefore, regularly conducting anti-icing inspections of transmission lines and achieving accurate visual identification and refined edge segmentation of iced insulators have become core prerequisites for current icing hazard monitoring and quantitative condition assessment.

[0003] Currently, using drones for power transmission line inspection has become an important method in power transmission line inspection. Insulator icing is the most common safety hazard in high-voltage power transmission lines in cold and high-altitude areas. Accurately segmenting ice-covered insulators is a key prerequisite for subsequent icing hazard assessment and fault early warning. Image processing-based methods for segmenting iced insulators are mainly divided into traditional image methods and deep learning methods. Traditional image processing methods: These methods typically use grayscale conversion, threshold segmentation, or manual feature extraction to extract insulator regions. However, in actual working conditions, iced images often have severe background interference, such as fog condensation, snow background, light reflection, and shadow occlusion. This makes traditional methods less robust when dealing with iced insulators in complex environments, and it is difficult to achieve fine edge segmentation. Deep learning-based methods: Although current convolutional neural networks perform well in target localization, they still have limitations in semantic segmentation tasks for iced insulators. First, ordinary convolutional operations are limited by local receptive fields, making it difficult to capture long-distance features. They also have weak global modeling capabilities for long, strip-shaped, and blurred-edge iced insulators. Secondly, existing models are not very accurate in handling the loss of texture details caused by ice reflection. Moreover, most of them remain at the stage of qualitative identification of presence or absence, lacking fine segmentation for the specific target of ice-covered insulators, making it difficult for power grid operation and maintenance departments to obtain intuitive and accurate morphological evidence of insulator icing. Summary of the Invention

[0004] To address the shortcomings of existing technologies in identifying icing thickness of insulators on transmission lines, this invention provides a method, system, equipment, and medium for identifying icing thickness of insulators on transmission lines.

[0005] To achieve the above objectives, the present invention provides the following technical solution: A method for identifying the ice thickness of transmission line insulators includes the following steps: An image of an ice-covered insulator on a transmission line is input into a pre-trained semantic segmentation model for ice-covered insulator images to obtain a binary segmented image of the insulator. The semantic segmentation model uses an improved HRNet network as the encoder and an UpperHead structure as the decoder. Specifically, the improved HRNet network introduces an FCA module into the high-resolution and medium-resolution feature branches of the original HRNet network, and cascades a CoordAtt module at the output of the FCA module. In the low-resolution feature branch of the original HRNet network, the basic residual module is replaced with a visual state-space model VSSBlock, resulting in the improved HRNet network. The FCA module combines frequency domain analysis and channel attention to capture key information in the feature map. The CoordAtt module is used to embed location information into the feature representation. The VSSBlock utilizes a dual-branch architecture to extract global topological features of the elongated insulator. The outermost closed contour of the connected domain of the icy insulator in the binary segmentation image of the insulator is extracted using a topological boundary tracking algorithm to obtain a set of boundary points. An uniced boundary point set is obtained based on an image of an uniced insulator at the same location and orientation. The ice-covered area and the reference area are calculated based on the boundary point set and the uniced boundary point set, respectively. The ice-covered area is compared with the reference area to calculate the rate of change of the graphic area, which characterizes the degree of ice expansion in the two-dimensional cross-section of the insulator. The axial and circumferential ice thickness of the insulator string is quantified based on the rate of change of the graphic area.

[0006] Preferably, in the FCA module, the feature tensor output from the previous stage is used as input, and the feature tensor is divided along the channel dimension into... Each group is assigned a set 2D DCT frequency component index. The feature tensor after the index is assigned is weighted and compressed using DCT basis functions to obtain a compressed frequency vector. The frequency vector is concatenated to obtain a multispectral frequency vector. The multispectral frequency vector is input into a fully connected layer and a Sigmoid activation function is used to generate channel weights. The input feature tensor is weighted and calibrated according to the channel weights.

[0007] Preferably, the recalibrated feature tensor is input into the CoordAtt module, and each channel is encoded along the horizontal and vertical directions using one-dimensional global average pooling kernels of different sizes. The encoded features in the two directions are concatenated along the channel dimension, and the concatenated features are mapped to a low-dimensional embedding space using a shared convolutional transformation. After the dimensionality-reduced features are modulated by a nonlinear activation function, they are then restored to the original channel dimension through convolutional mapping. Attention weight coefficients along the height and width directions are generated respectively, and multiplied with the recalibrated feature tensor channel by channel to obtain the enhanced features.

[0008] Preferably, in the low-resolution feature branch, the feature tensor output from the previous stage is used as input for feature extraction to obtain a low-resolution feature tensor; the low-resolution feature tensor is input into the VSSBloc, first processed by LayerNorm, and then split into the main branch and the gated branch in the channel dimension. In the main branch, the low-resolution feature tensor is passed through a linear projection layer, a deep convolutional layer, and a SiLU activation function in sequence to capture local spatial features. The local spatial features are then input into the 2D selective scanning module for global feature modeling, and the resulting global features are normalized again using LayerNorm. The gating branch processes the low-resolution feature tensor through a linear layer and a SiLU activation function to generate an adjustment signal. The normalized global features and the modulated signal are multiplied element-wise for feature fusion, channel mixing is performed through the linear layer at the end, and residual connections are performed with the low-resolution feature tensor to output global topological features.

[0009] Preferably, the training of the semantic segmentation model for icing insulator images also includes acquiring images of icing insulators on transmission lines through a fixed-position image video monitoring terminal deployed on the transmission tower; using a preset annotation tool to annotate the icing insulators in the images to obtain semantic segmentation labels; obtaining an icing insulator dataset based on the labeled images of icing insulators on transmission lines; and dividing the dataset into a training set and a validation set. The training set is augmented using a hybrid dehybrid MUM data augmentation strategy. The augmented training set data is then used to train the semantic segmentation model for iced insulator images, resulting in a pre-trained semantic segmentation model for iced insulator images.

[0010] Preferably, the ice-covered area and the reference area of ​​the un-ice-covered insulator are calculated based on the boundary point set and the un-ice-covered boundary point set, respectively. Specifically, the extracted contour is first approximated as an m-sided shape based on the boundary point set. The area of ​​the region enclosed by the outline is calculated using Green's formula, specifically using the following formula: ; in, The area enclosed by the closed polygonal contour, including the area enclosed by ice and the base area enclosed by no ice. The total number of vertices of the approximately extracted polygonal contour; , ) is the first polygonal outline Two-dimensional pixel coordinates of each vertex; For the closed contour adjacent to the first The two-dimensional pixel coordinates of the next vertex of the given vertex, and it is specified that when At that time, there exists a closed condition where the beginning and end are connected. ; Indicates the connection of the first The vertex and the first The boundary of a directed line segment with vertices.

[0011] Preferably, the area enclosed by the ice is compared with the reference enclosed area to calculate the rate of change of the graphic area, which characterizes the degree of ice expansion in the two-dimensional cross-section of the insulator. The rate of change of the graphic area is specifically calculated using the following formula: ; in, The baseline area enclosed by the un-ice-covered area; This refers to the area surrounded by ice.

[0012] This invention also provides a system for identifying the icing thickness of transmission line insulators, specifically including: An image processing module is used to acquire images of ice-covered insulators on transmission lines and input them into a pre-trained semantic segmentation model for ice-covered insulator images to obtain binary segmented images of the insulators. The semantic segmentation model for ice-covered insulator images uses an improved HRNet network as the encoder and an UpperHead structure as the decoder. Specifically, the improved HRNet network introduces an FCA module into the high-resolution and medium-resolution feature branches of the original HRNet network, and cascades a CoordAtt module at the output of the FCA module. In the low-resolution feature branch of the original HRNet network, the basic residual module is replaced with the visual state space model VSSBlock to obtain the improved HRNet network. The FCA module combines frequency domain analysis and channel attention to capture key information in the feature map. The CoordAtt module is used to embed position information into the feature representation. The VSSBlock uses a dual-branch architecture to extract the global topological features of the elongated insulators.

[0013] The thickness recognition module is used to extract the outermost closed contour of the connected domain of the icy insulator in the binary segmentation image of the insulator using a topological boundary tracking algorithm, and obtain a set of boundary points. Based on the image of the uniced insulator at the same position and in the same posture, an uniced boundary point set is obtained. Based on the set of boundary points and the uniced boundary point set, the ice-covered area and the reference area are calculated respectively. The ice-covered area is compared with the reference area to calculate the graphic area change rate, which characterizes the degree of ice expansion of the current two-dimensional cross-section of the insulator. Based on the graphic area change rate, the axial and circumferential ice thickness of the insulator string is quantified.

[0014] The present invention also provides a computer device, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps described in the method for identifying the ice thickness of transmission line insulators.

[0015] The present invention also provides a computer-readable storage medium storing a computer program that, when loaded by a processor, can execute the steps described in the method for identifying the ice thickness of transmission line insulators.

[0016] The method for identifying the ice thickness of transmission line insulators provided by this invention has the following beneficial effects: This invention introduces VSSBlock into the original HRNet segmentation network, effectively overcoming the limitations of the local receptive field of traditional convolutions. While maintaining linear computational complexity, it significantly improves the ability to extract the overall contour of long, strip-shaped insulators. By cascading FCA and CoordAtt in the residual module, it fully integrates complex textures in the frequency domain and precise spatial location information, significantly enhancing the model's detection accuracy for minute edge details against reflective ice and snow backgrounds. Based on the binary segmented image of the insulator obtained from the model processing, the outer closed contour is extracted using edge operators. The double area integral is simplified to the algebraic summation of shoelaces on discrete polygons, achieving a dual improvement in computational speed and noise resistance stability. This enables efficient and high-precision quantitative identification of the icing state of transmission line insulators, providing reliable support for intelligent anti-icing inspection of the power grid. Attached Figure Description

[0017] To more clearly illustrate the embodiments and design schemes of the present invention, the accompanying drawings required for this embodiment will be briefly described below. The drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart of the method for identifying the ice thickness of transmission line insulators according to an embodiment of the present invention.

[0019] Figure 2 This is a schematic diagram of the overall architecture of the improved HRNet segmentation network in an embodiment of the present invention.

[0020] Figure 3 This is a schematic diagram of the frequency channel attention mechanism in an embodiment of the present invention.

[0021] Figure 4 This is a schematic diagram of the coordinate attention mechanism structure in an embodiment of the present invention.

[0022] Figure 5 This is a schematic diagram of the 2D selective scanning mechanism in an embodiment of the present invention.

[0023] Figure 6 The images show the segmentation effects of HRNet on composite insulators before and after the improvement in this embodiment of the invention. Figure 6 of (a) Figure 6 (b) and Figure 6 (c) shows the segmentation effect diagrams for different scenarios. Detailed Implementation

[0024] To enable those skilled in the art to better understand and implement the technical solutions of the present invention, the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. The following embodiments are only used to more clearly illustrate the technical solutions of the present invention and should not be construed as limiting the scope of protection of the present invention.

[0025] Example This invention provides a method for identifying the ice thickness of insulators on power transmission lines, such as... Figure 1 As shown, the specific steps include: S1. Obtain images of ice-covered insulators on transmission lines through fixed-position image and video monitoring terminals deployed on transmission towers.

[0026] Using a pre-defined annotation tool, ice-covered insulators in the image are labeled to obtain semantic segmentation labels. Manual annotation is then performed using the Labelme open-source image annotation tool. The Labelme tool is used to delineate the outermost physical contour of the ice-covered insulator point-by-point with polygons, accurately identifying the ice-covered insulator as a complete recognition target. After annotation, the annotation information of the ice-covered insulators in a single frame image is saved as a JSON file and further batch-converted into binary mask images for use in training the semantic segmentation network. The insulator annotation information mainly includes: category labels for the ice-covered insulator and the background, as well as the set of vertex coordinates of the discrete polygons constituting the closed contour of the ice-covered insulator.

[0027] S2. Preprocess the sample data to construct an ice-covered insulator dataset for training and performance verification of the deep learning semantic segmentation model, and divide the dataset into a training set and a validation set.

[0028] S3. To address the issue of a small number of original samples, a hybrid demixing MUM data augmentation strategy is adopted to augment the training set, enriching the diversity of training samples while preserving the spatial geometric positioning information of the ice-covered insulators.

[0029] The MUM data augmentation strategy works as follows: In each training batch, four images are randomly selected and evenly divided into 16 image patches in the spatial dimension. These patches are then cross-stitched according to a randomly generated spatial permutation matrix to generate a hybrid image with extremely high scene complexity. Simultaneously, this hybrid image is stored in the spatial location index record of this patch exchange. Subsequently, the synthesized hybrid image is input into the segmentation network for forward propagation, forcing the network to extract robust local icing texture features in highly fragmented and strongly interfered images. Before the feature map is passed to the decoder after the encoding stage, the pre-saved spatial location index record is extracted for reverse routing. This decouples, peels off, and repositions the feature patches in the hybrid feature map, restoring them to the reconstructed feature map corresponding to the initial spatial topology of the original and target images.

[0030] S4. Construct a semantic segmentation model for iced insulator images based on HRNet. The constructed semantic segmentation model for iced insulator images includes an encoder structure and a decoder structure. The encoder structure adopts an improved HRNet network; the decoder adopts an UperHead structure for pyramid pooling and multi-level fusion of the multi-scale features output by the encoder. The semantic segmentation model for iced insulator images is as follows: Figure 2 As shown.

[0031] The original HRNet network consists of four stages: the first stage includes the Bottleneck module; the second stage includes high-resolution and medium-resolution feature branches; the third stage includes high-resolution, medium-resolution, and low-resolution feature branches; and the fourth stage includes high-resolution, medium-resolution, and two low-resolution feature branches. In the low-resolution deep branches of the original HRNet network, the original basic residual module is replaced with a VSSBlock module; in the high-resolution and medium-resolution feature branches of the original HRNet network, an FCA module is introduced, and a CoordAtt module is cascaded at the output of the FCA module to obtain the improved HRNet network.

[0032] Frequency Channel Attention (FCA) mechanism, such as Figure 3 As shown, its implementation process is as follows: given an input feature tensor First, it is divided along the channel dimension into There are groups, each group is represented as ,in For each group, a specific 2D DCT frequency component index is assigned. We use DCT basis functions to perform weighted compression on the features. The frequency components can be expressed as: ; in, These are DCT basis functions. The compressed frequency vectors for each group. After concatenation, a complete multispectral frequency vector is formed. Subsequently, this vector is mapped through a fully connected layer and used with a sigmoid activation function to generate channel weights. And recalibrate the original input. .

[0033] CoordAtt mechanism, such as Figure 4 As shown, the implementation process is as follows: for the recalibration feature output by the FCA module... Use sizes respectively and A one-dimensional global average pooling kernel, along the horizontal coordinate and vertical coordinates Encode each channel. Each channel is at a height Output at With in width Output at They can be represented as: ; ; Will and The encoded features from both directions are concatenated along the channel dimension and mapped to a low-dimensional embedding space via a shared convolutional transformation, completing cross-directional information interaction and correlation modeling in the compressed representation. After modulation by a nonlinear activation function, the original channel dimension is restored through convolutional mapping, generating attention weight coefficients along the height and width directions respectively, which are then multiplied with the input features channel by channel to achieve reweighting. This mechanism preserves the channel attention modeling capability while embedding explicit positional information into the feature representation process, making the spatial response directionally sensitive and coordinate-dependent, thereby enhancing the network's spatial perception capability of the slender structure and topological extension features of insulators.

[0034] The Visual State Space Model (VSSBlock) employs a two-branch architecture based on layer normalization. The input feature tensor is first processed by LayerNorm, then split along the channel dimension into the main processing branch and the gate branch. In the main branch, features are sequentially passed through a linear projection layer, a depth-wise convolution layer with a kernel size of , and a SiLU activation function to capture local spatial features. These features are then fed into the 2D Selective Scan (SS2D) module for global feature modeling, and are normalized again by LayerNorm before output. The gate branch, consisting of a linear layer and a SiLU activation function, generates modulation signals. The outputs of both branches undergo element-wise multiplication for feature fusion, and finally, channel mixing is performed through the final linear layer. A residual connection is then established with the original input to output the final features.

[0035] 2D selective scanning mechanism (SS2D) such as Figure 5 As shown, its implementation process includes three steps: cross-scan, selective scanning using the S6 module, and cross-merging. The input feature map first undergoes a cross-scan operation, flattening and recombining the 2D feature map into four 1D sequences of length 1 along four directions: top-left to bottom-right, bottom-right to top-left, top-right to bottom-left, and bottom-left to top-right. These sequences are input in parallel into the S6 block, where a forward scan operation is performed using the discretized input dependency parameters. After processing, the four sequences are reverse-engineered back to their original 2D spatial arrangement and then summed and merged through a cross-merge operation, thus achieving the aggregation of full-map context information while maintaining linear memory and computational overhead.

[0036] The image of an ice-covered insulator is input into the semantic segmentation model for ice-covered insulator images. First, the image is resized to a standard-size tensor. Then, an improved HRNet encoder is used for multi-resolution feature extraction. Specifically, the low-resolution deep branch uses VSSBlock to extract global topological features of the elongated insulators; the high- and medium-resolution branches pass through the FCA and CoordAtt modules sequentially to enhance the perception of ice and snow textures and edge details. Subsequently, the features from each branch undergo multi-scale interaction and fusion. The fused features are then input into the UpperHead decoder, where a pyramid pooling module (PPM) aggregates contextual information and performs feature concatenation to restore spatial resolution. The decoded feature map is then subjected to pixel-level prediction by a classification convolutional layer, ultimately outputting a binary segmented image of the insulator with the same resolution as the original image.

[0037] S5. Using the ice-covered insulator data training set constructed in S1, perform end-to-end weight update and convergence training on the improved HRNet segmentation network.

[0038] S6. Using the trained improved HRNet segmentation network and insulator weight data, test the ice-covered insulator data validation set to obtain pixel-level segmentation results of the insulator region and generate binary segmented images of the insulator. The segmentation results of composite insulators using HRNet before and after the improvement are shown in the following figures. Figure 6 As shown.

[0039] S7. Using the topological boundary tracking algorithm, extract the outermost continuous closed contour boundary point set of the connected domain of the ice-covered insulator from the segmented binary image.

[0040] S8. Approximate the extracted contour as an m-sided polygon and use Green's formula to calculate the area of ​​the region enclosed by the contour.

[0041] By utilizing a special case of Stokes' theorem in a two-dimensional plane, the double area integral of a region is transformed into a one-dimensional line integral along a closed contour, and finally simplified to an algebraic summation formula for the vertex coordinates of a discrete polygon (Shoelace Formula). The area calculation formula is as follows: ; in, The area of ​​the region enclosed by the closed contour of the polygon; The total number of vertices of the approximately extracted polygonal contour; , ) is the first polygonal outline Two-dimensional pixel coordinates of each vertex; For the closed contour adjacent to the first The two-dimensional pixel coordinates of the next vertex of the given vertex, and it is specified that when At that time, there exists a closed condition where the beginning and end are connected. ; Indicates the connection of the first The vertex and the first The boundary of a directed line segment with vertices.

[0042] S9. Obtain historical images of the same insulator in its un-iced state, and calculate its outline area using the same method. By subtracting and scaling the absolute geometric areas of the insulator before and after icing, calculate the rate of change of the graphic area that characterizes the degree of icing expansion of the current two-dimensional cross-section of the insulator.

[0043] Specifically, the actual connected region geometric area of ​​the image of the ice-covered insulator is accurately calculated using Green's theorem in step S8. Then, the system will retrieve the pre-stored reference area of ​​the insulators that are not covered by ice from the same viewing angle. This invention calculates the rate of change of the graphic area by subtracting the absolute geometric areas of the two and scaling them out. The calculation formula is as follows: ; This rate of change can intuitively and accurately reflect the degree of morphological expansion of the insulator in a two-dimensional cross-section caused by icing, and comprehensively evaluate its axial and circumferential icing thickness to quantitatively analyze the degree of change in the insulator morphology with the increase or decrease of icing.

[0044] The MUM data augmentation technique overcomes the bottleneck of small sample sizes. The improved HRNet segmentation network integrates global state-space modeling and a frequency-domain-position attention mechanism, significantly enhancing edge segmentation accuracy under complex weather conditions. Rigorous Green's formula is incorporated for geometric topology calculations, and physical quantization is performed based on the segmentation mask, resulting in extremely high theoretical confidence and noise resistance. End-to-end recognition is achieved from UAV inspection images to the final high-precision equivalent ice thickness detection result, demonstrating high accuracy and robustness, making it suitable for various ice disaster prevention and inspection scenarios.

[0045] This invention also provides a system for identifying the icing thickness of transmission line insulators, comprising: The image processing module is used to acquire images of iced insulators on transmission lines. The pre-trained semantic segmentation model for iced insulator images is then used to obtain binary segmented images of the insulators. The semantic segmentation model uses an improved HRNet network as the encoder and an UpperHead structure as the decoder. Specifically, the improved HRNet network introduces an FCA module into the high-resolution and medium-resolution feature branches of the original HRNet network, and cascades a CoordAtt module at the output of the FCA module. In the low-resolution feature branch of the original HRNet network, the basic residual module is replaced with the visual state space model VSSBlock, resulting in the improved HRNet network. The FCA module combines frequency domain analysis and channel attention to capture key information in the feature map. The CoordAtt module embeds location information into the feature representation. VSSBlock utilizes a dual-branch architecture to extract the global topological features of the elongated insulators.

[0046] The thickness recognition module is used to extract the outermost closed contour of the connected domain of the icy insulator in the binary segmentation image of the insulator using a topological boundary tracking algorithm, and obtain the boundary point set. The uniced boundary point set is obtained based on the image of the uniced insulator at the same position and in the same posture. The ice-covered area and the reference area are calculated based on the boundary point set and the uniced boundary point set, respectively. The ice-covered area is compared with the reference area to calculate the graphic area change rate, which represents the degree of ice expansion of the current two-dimensional cross-section of the insulator. The axial and circumferential ice thickness of the insulator string is obtained based on the graphic area change rate.

[0047] Each module in the aforementioned transmission line insulator icing thickness identification system can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the computer device's memory as software, so that the processor can call and execute the corresponding operations of each module.

[0048] The present invention also provides a computer device, including a memory, a processor, and a computer program stored in the memory. The processor executes the computer program to implement the steps in the embodiment of the method for identifying the ice thickness of transmission line insulators. Specific implementation methods can be found in the method embodiments, and will not be repeated here.

[0049] Furthermore, the present invention also provides a non-transitory computer-readable storage medium containing instructions, on which a computer program is stored. For example, a memory containing instructions that can be executed by a processor of a computer device to perform the above-described method. For example, the non-transitory computer-readable storage medium may be a ROM, random access memory (RAM), CD-ROM, magnetic tape, floppy disk, and optical data storage device, etc. When the computer program is executed by the processor, it can implement the steps in the embodiments of the method for identifying the ice thickness of transmission line insulators. Specific implementation methods can be found in the method embodiments, which will not be repeated here.

[0050] Those skilled in the art will understand that embodiments of the present invention can provide methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0051] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, as well as combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0052] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0053] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0054] It should be noted that the specific embodiments described above enable those skilled in the art to more fully understand the present invention, but do not limit the present invention in any way. Therefore, although the present invention has been described in detail in this specification and embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the present invention; and all technical solutions and improvements that do not depart from the spirit and scope of the present invention are covered within the protection scope of the present invention patent. No reference numerals in the claims should be construed as limiting the scope of the claims. Any simple variations or equivalent substitutions of technical solutions that can be readily obtained by those skilled in the art within the scope of the technology disclosed in the present invention are within the protection scope of the present invention.

Claims

1. A method for identifying the ice thickness of insulators on transmission lines, characterized in that, Includes the following steps: An image of an ice-covered insulator on a transmission line is input into a pre-trained semantic segmentation model for ice-covered insulator images to obtain a binary segmented image of the insulator. The semantic segmentation model uses an improved HRNet network as the encoder and an UpperHead structure as the decoder. Specifically, the improved HRNet network introduces an FCA module into the high-resolution and medium-resolution feature branches of the original HRNet network, and cascades a CoordAtt module at the output of the FCA module. In the low-resolution feature branch of the original HRNet network, the basic residual module is replaced with a visual state-space model VSSBlock, resulting in the improved HRNet network. The FCA module combines frequency domain analysis and channel attention to capture key information in the feature map. The CoordAtt module is used to embed location information into the feature representation. The VSSBlock utilizes a dual-branch architecture to extract global topological features of the elongated insulator. The outermost closed contour of the connected domain of the icy insulator in the binary segmentation image of the insulator is extracted using a topological boundary tracking algorithm to obtain a set of boundary points. An uniced boundary point set is obtained based on an image of an uniced insulator at the same location and orientation. The ice-covered area and the reference area are calculated based on the boundary point set and the uniced boundary point set, respectively. The ice-covered area is compared with the reference area to calculate the rate of change of the graphic area, which characterizes the degree of ice expansion in the two-dimensional cross-section of the insulator. The axial and circumferential ice thickness of the insulator string is quantified based on the rate of change of the graphic area.

2. The method for identifying the ice thickness of transmission line insulators according to claim 1, characterized in that, In the FCA module, the feature tensor output from the previous stage is used as input, and the feature tensor is divided along the channel dimension. Each group is assigned a set 2D DCT frequency component index. The feature tensor after the index is assigned is weighted and compressed using DCT basis functions to obtain a compressed frequency vector. The frequency vector is concatenated to obtain a multispectral frequency vector. The multispectral frequency vector is input into a fully connected layer and a Sigmoid activation function is used to generate channel weights. The input feature tensor is weighted and calibrated according to the channel weights.

3. The method for identifying the ice thickness of transmission line insulators according to claim 2, characterized in that, The recalibrated feature tensor is input into the CoordAtt module. Using one-dimensional global average pooling kernels of different sizes, each channel is encoded along the horizontal and vertical directions. The encoded features in the two directions are concatenated along the channel dimension. The concatenated features are mapped to a low-dimensional embedding space using a shared convolutional transformation. After the dimensionality-reduced features are modulated by a non-linear activation function, they are restored to the original channel dimension through convolutional mapping. Attention weight coefficients along the height and width directions are generated respectively. These are multiplied channel by channel with the recalibrated feature tensor to obtain the enhanced features.

4. The method for identifying the ice thickness of transmission line insulators according to claim 1, characterized in that, In the low-resolution feature branch, the feature tensor output from the previous stage is used as input for feature extraction to obtain a low-resolution feature tensor. The low-resolution feature tensor is then input into the VSSBloc, processed by LayerNorm, and split into the main branch and the gated branch in the channel dimension. In the main branch, the low-resolution feature tensor is passed through a linear projection layer, a deep convolutional layer, and a SiLU activation function in sequence to capture local spatial features. The local spatial features are then input into the 2D selective scanning module for global feature modeling, and the resulting global features are normalized again using LayerNorm. The gating branch processes the low-resolution feature tensor through a linear layer and a SiLU activation function to generate an adjustment signal. The normalized global features and the modulated signal are multiplied element-wise for feature fusion, channel mixing is performed through the linear layer at the end, and residual connections are performed with the low-resolution feature tensor to output global topological features.

5. The method for identifying the ice thickness of transmission line insulators according to claim 1, characterized in that, It also includes training the semantic segmentation model for icing insulator images. Specifically, it involves acquiring images of icing insulators on transmission lines through fixed-position image and video monitoring terminals deployed on transmission towers; using a pre-set annotation tool to annotate the icing insulators in the images to obtain semantic segmentation labels; obtaining an icing insulator dataset based on the labeled images of icing insulators on transmission lines; and dividing the dataset into a training set and a validation set. The training set is augmented using a hybrid dehybrid MUM data augmentation strategy. The augmented training set data is then used to train the semantic segmentation model for iced insulator images, resulting in a pre-trained semantic segmentation model for iced insulator images.

6. The method for identifying the ice thickness of transmission line insulators according to claim 1, characterized in that, Based on the boundary point set and the un-iced boundary point set, the ice-covered area and the reference area of ​​the un-iced insulator are calculated respectively. Specifically, the extracted contour is first approximated as an m-sided shape based on the boundary point set. The area of ​​the region enclosed by the outline is calculated using Green's formula, specifically using the following formula: ； in, The area enclosed by the closed polygonal contour, including the area enclosed by ice and the base area enclosed by no ice. The total number of vertices of the approximately extracted polygonal contour; , ) is the first polygonal outline Two-dimensional pixel coordinates of each vertex; For the closed contour adjacent to the first The two-dimensional pixel coordinates of the next vertex of the given vertex, and it is specified that when At that time, there exists a closed condition where the beginning and end are connected. ; Indicates the connection of the first The vertex and the first The boundary of a directed line segment with vertices.

7. The method for identifying the ice thickness of transmission line insulators according to claim 6, characterized in that, By comparing the area enclosed by the ice accretion with the reference enclosed area, the rate of change of the graphic area, which characterizes the degree of ice expansion in the two-dimensional cross-section of the insulator, is calculated. The rate of change of the graphic area is specifically calculated using the following formula: ； in, The baseline area enclosed by the un-ice-covered area; The area surrounded by ice.

8. A system for identifying the ice thickness of insulators on transmission lines, characterized in that, include: The image processing module is used to obtain images of ice-covered insulators on transmission lines, input them into a pre-trained semantic segmentation model for ice-covered insulator images, and obtain binary segmented images of the insulators. The semantic segmentation model for iced insulator images uses an improved HRNet network as the encoder and an UpperHead structure as the decoder. Specifically, the improved HRNet network introduces an FCA module into the high-resolution and medium-resolution feature branches of the original HRNet network, and cascades a CoordAtt module at the output of the FCA module. In the low-resolution feature branch of the original HRNet network, the basic residual module is replaced with the visual state-space model VSSBlock, resulting in the improved HRNet network. The FCA module combines frequency domain analysis and channel attention to capture key information in the feature map; the CoordAtt module embeds positional information into the feature representation; and the VSSBlock utilizes a dual-branch architecture to extract the global topological features of the elongated insulator. The thickness recognition module is used to extract the outermost closed contour of the connected domain of the icy insulator in the binary segmentation image of the insulator using a topological boundary tracking algorithm, and obtain a set of boundary points. Based on the image of the uniced insulator at the same position and in the same posture, an uniced boundary point set is obtained. Based on the set of boundary points and the uniced boundary point set, the ice-covered area and the reference area are calculated respectively. The ice-covered area is compared with the reference area to calculate the graphic area change rate, which characterizes the degree of ice expansion of the current two-dimensional cross-section of the insulator. Based on the graphic area change rate, the axial and circumferential ice thickness of the insulator string is quantified.

9. A computer device, comprising a memory, a processor, and a computer program stored in the memory, characterized in that, The processor executes the computer program to implement the steps of the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is loaded by the processor, it is able to perform the steps of the method according to any one of claims 1 to 7.

Images