Power transmission line double-view image icing monitoring method and device

Abstract

The invention relates to the technical field of icing monitoring, in particular to a power transmission line double-view image icing monitoring method and device, based on a double-view icing image, more icing details are obtained in a complex meteorological scene through multi-view information complementation, the accuracy of icing type judgment is effectively improved, and the accuracy of icing type judgment is improved. The accuracy of ice coating density calculation is improved; moreover, based on a lead standard ice thickness extraction technology of a double-view-angle image, icing thicknesses in two mutually perpendicular directions can be obtained, the accuracy of icing sectional area calculation is effectively improved by making lead icing with an irregular section approximate to an ellipse, and the lead icing is converted into standard icing with a circular section in combination with an equal-weight method, so that the accuracy of icing sectional area calculation is improved. And finally, the accuracy of standard icing thickness calculation is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of ice-coating monitoring, in particular to a transmission line double-view image ice-coating monitoring method and device. BACKGROUND

[0002] Under the conditions of low temperature, high humidity, wind or precipitation, supercooled water droplets, raindrops, fog droplets or wet snow in the air freeze rapidly after contacting the surface of the transmission conductor with a temperature lower than 0℃, forming an ice layer on the outer surface of the conductor, which is called ice-coating. Common types of ice-coating include rime, glaze, mixed rime and wet snow adhesion. Under the background of global warming, the frequency and intensity of ice-coating on transmission lines have increased significantly.

[0003] The distance between the farthest two points on the surface of the ice layer on the cross section of the conductor is called the ice-coating thickness. As the ice-coating thickness of the conductor increases, it will cause excessive conductor sag and insufficient distance from surrounding grounded objects, resulting in flashover. At the same time, the vertical, horizontal and longitudinal loads on the tower increase rapidly, and when the combined load exceeds the design load of the tower, the tower will be damaged or collapse, seriously affecting the safe operation of the power grid. Timely, accurate and effective ice-coating monitoring means is the prerequisite for ice damage prevention work. Compared with manual observation, online ice-coating monitoring has the advantages of small workload and good continuity, mainly including image method, tension sensing method and optical fiber sensing method. Based on images, the ice-coating process, shape, degree and other information can be directly obtained, which has the advantages of intuitive results, rich identification information and rapid research and judgment support. The ice-coating image online monitoring device has been widely applied in power grids.

[0004] The load generated by the ice-coating per unit length of the conductor is related to the cross-sectional area and density of the ice-coating. The actual ice-coating of the conductor has different cross-sectional shapes and uneven thickness, and the density varies greatly depending on the type of ice-coating. The maximum ice-coating value that the line can bear is specified during line design, which is called design ice thickness. The design ice thickness considers the ideal case where the ice-coating cross section is circular, the thickness is uniform, and the density is 0.9 g / cm 3 . In order to compare with the design ice thickness, evaluate the operation state of the line and decide whether to start ice melting, the actual ice-coating is usually approximated as an ellipse, and then the cross section is converted to a circle by the equal weight method, and the standard ice thickness is calculated according to the ice-coating density of 0.9 g / cm 3 . The principle of standard ice thickness calculation is shown in Figure 1 .

[0005] Referring to Figure 2As shown in the prior art, the current image-based online icing monitoring method mainly relies on single-view images to identify the icing condition, obtains the distance between the farthest two points on the icing surface in one direction of the conductor cross-section by comparing the pixel changes of the conductor width without icing and with icing, and calculates the icing thickness. However, due to the limited view angle of the monocular camera, only calculating the icing thickness in one direction cannot effectively calculate the icing cross-sectional area, which further affects the calculation of the standard icing thickness.

[0006] In the disclosed patent document CN119169535A, it is proposed to use a monocular camera installed on the conductor to identify the front-view icing thickness and the side-view icing thickness of the main-view conductor and the side-view conductor respectively, and then use the equal weight method to approximately calculate the circular cross-sectional area and further calculate the standard icing thickness. The main problems of this scheme are: first, some lines do not have side-view conductors, so the side-view icing thickness cannot be obtained; second, the image monitoring device is installed on the main-view conductor, and the spatial relative position between the side-view conductor and the main-view conductor (icing monitoring device) will change after icing, so the comparison of the pixel positions of the side-view conductor before and after icing through image semantic segmentation will produce deviation, resulting in errors in the calculation of the icing thickness.

[0007] Moreover, the existing technology mainly relies on the brightness and texture features of the icing image to first identify the icing type, and then assigns values to different categories of pixels in the image based on the preset density recommendation values of different icing types. However, the icing process is often accompanied by complex weather conditions (strong light, heavy fog, etc.), and single-view images are easily disturbed by heavy fog, mirror reflection, and local occlusion, leading to "key features invisible", affecting the identification of icing brightness and texture features, and easily causing icing type misjudgment, which affects the calculation results of icing density. The icing recognition effect of the single-view image is shown in Figure 3 .

[0008] In summary, the method of calculating the icing cross-sectional area and the icing density based on single-view images in the existing technology, especially in complex weather environments, easily leads to large errors in the calculation of the standard icing thickness. SUMMARY

[0009] Therefore, the technical problem to be solved by the present application is to overcome the low accuracy of the conductor standard icing thickness calculation method based on single-view images in the existing technology.

[0010] To solve the above technical problems, the present application provides a double-view image icing monitoring method for a power transmission line, comprising: obtaining a front-view original icing image and a side-view original icing image of the power transmission line conductor collected synchronously; performing gray scale conversion and histogram equalization on the front-view original icing image and the side-view original icing image to obtain a front-view pre-processed icing image and a side-view pre-processed icing image; The feature extraction is respectively performed on the icing image after the front view pretreatment and the icing image after the side view pretreatment, and the front view high-level semantic feature and the side view high-level semantic feature are correspondingly obtained; the front view high-level semantic feature and the side view high-level semantic feature are spliced and sequentially subjected to the channel attention mechanism and the spatial attention mechanism, and the shared encoding feature is obtained; The shared encoding feature, the front view high-level semantic feature and the icing image after the front view pretreatment are input into the semantic segmentation network, and the front view segmentation image is obtained; the shared encoding feature, the side view high-level semantic feature and the icing image after the side view pretreatment are input into the semantic segmentation network, and the side view segmentation image is obtained; The segmentation results of the front view segmentation image and the side view segmentation image are screened and optimized by taking the proportion of the effective pixels in the segmentation image and the proportion of the highlight pixels in the original icing image as the judgment indexes, and the pixel proportion of each icing type in the front view and the pixel proportion of each icing type in the side view are obtained; The icing density is calculated based on the pixel proportion of each icing type in the front view and the pixel proportion of each icing type in the side view; the front view icing thickness is calculated based on the front view segmentation image; the side view icing thickness is calculated based on the side view segmentation image; and the standard icing thickness is calculated based on the icing density, the front view icing thickness and the side view icing thickness.

[0011] Preferably, the gray conversion and the histogram equalization are performed on the front view original icing image and the side view original icing image to obtain the icing image after the front view pretreatment and the icing image after the side view pretreatment, and the steps include: The gray conversion is performed on the front view original icing image and the side view original icing image to obtain the front view gray icing image and the side view gray icing image; The conductor icing area of the front view gray icing image and the side view gray icing image is respectively obtained, and the conductor icing area is divided into a plurality of non-overlapping sub-blocks; The gray histogram of each sub-block is calculated; a preset clipping threshold is selected based on the current meteorological scene, the pixels exceeding the clipping threshold in each sub-block are clipped, and the number of the clipped pixels in each sub-block is counted; the number of the clipped pixels in each sub-block is evenly divided into all the gray levels of the sub-block, and the enhanced gray histogram of the sub-block is obtained; Based on the enhanced gray histogram of each sub-block, each pixel value of the sub-block is mapped to a new gray value, and the icing image after the front view pretreatment and the icing image after the side view pretreatment are obtained.

[0012] Preferably, the feature extraction is respectively performed on the icing image after the front view pretreatment and the icing image after the side view pretreatment, and the front view high-level semantic feature and the side view high-level semantic feature are correspondingly obtained, and the method is as follows: The ice-coated image preprocessed at the front view angle is input into a first feature extractor to obtain a high-level semantic feature at the front view angle; the high-level semantic feature at the front view angle comprises a front view angle fog feature and a front view angle illumination feature; The ice-coated image preprocessed at the side view angle is input into a second feature extractor to obtain a high-level semantic feature at the side view angle; the high-level semantic feature at the side view angle comprises a side view angle fog feature and a side view angle illumination feature.

[0013] Preferably, the first feature extractor and the second feature extractor adopt DeepLabv3+, ResNet50 or MobileNetV3.

[0014] Preferably, the high-level semantic feature at the front view angle and the high-level semantic feature at the side view angle are spliced and then sequentially subjected to a channel attention mechanism and a spatial attention mechanism to obtain a shared encoding feature, and the steps comprise: The high-level semantic feature at the front view angle and the high-level semantic feature at the side view angle are spliced to obtain a spliced feature tensor; The spliced feature tensor is subjected to the channel attention mechanism to obtain a channel attention feature; The channel attention feature is subjected to the spatial attention mechanism to obtain a shared feature; The shared feature and the spliced feature tensor are added and then subjected to a convolution layer to obtain the shared encoding feature.

[0015] Preferably, the spliced feature tensor is subjected to the channel attention mechanism to obtain the channel attention feature, and the steps comprise: The spliced feature tensor is subjected to a first convolution layer, a second convolution layer, a global average pooling layer, a first full connection layer, a ReLU activation function, a second full connection layer and a Sigmoid activation function connected in sequence to obtain a channel attention weight; The channel attention weight and the spliced feature tensor are subjected to a channel-by-channel element multiplication operation to obtain the channel attention feature.

[0016] Preferably, the channel attention feature is subjected to the spatial attention mechanism to obtain the shared feature, and the steps comprise: The channel attention feature is respectively subjected to a global average pooling layer and a global maximum pooling layer and then spliced to obtain a fusion feature; The fusion feature is subjected to a third convolution layer and a Sigmoid activation function connected in sequence to obtain a spatial attention weight; The spatial attention weight and pixel values of corresponding spatial positions of the channel attention feature are multiplied one by one to obtain the shared feature.

[0017] Preferably, the semantic segmentation network adopts DeepLabv3+.

[0018] Preferably, the shared coding features, frontal view high-level semantic features, and the frontal view preprocessed icing image are input into the semantic segmentation network to obtain the frontal view segmented image. The steps include: The preprocessed icy image from the frontal view is input into the encoder of the semantic segmentation network to obtain the frontal view encoded features. The frontal view encoded features are then input into the hollow spatial pyramid pooling module to obtain the frontal view intermediate features. The shared encoded features are input into the hollow spatial pyramid pooling module to obtain the shared intermediate features. The frontal view intermediate features and the shared intermediate features are concatenated to obtain the frontal view context features. The frontal view context features are then concatenated with the frontal view high-level semantic features and passed sequentially through a Dropout layer, a normalization layer, and a ReLU activation function before being input into the decoder of the semantic segmentation network to obtain the frontal view segmented image. The shared coding features, high-level semantic features from the side view, and the preprocessed icing image from the side view are input into the semantic segmentation network to obtain the side view segmented image. The steps include: The preprocessed icy image from the side view is input into the encoder of the semantic segmentation network to obtain the side view encoded features. The side view encoded features are then input into the hollow spatial pyramid pooling module to obtain the side view intermediate features. The shared encoded features are input into the hollow spatial pyramid pooling module to obtain the shared intermediate features. The side view intermediate features and the shared intermediate features are concatenated to obtain the side view context features. The side view context features and the side view high-level semantic features are then concatenated and passed through a Dropout layer, a normalization layer, and a ReLU activation function, and then input into the decoder of the semantic segmentation network to obtain the side view segmented image.

[0019] Preferably, the segmentation results of the frontal view segmentation image and the side view segmentation image are filtered and optimized using the proportion of effective pixels in the segmented image and the proportion of bright pixels in the original icing image as judgment indicators, to obtain the pixel proportions of each icing type in the frontal view and the pixel proportions of each icing type in the side view. The steps include: Calculate the percentage of effective pixels in a frontal view segmented image. The percentage of effective pixels in a side-view segmented image ; Calculate the proportion of bright pixels in the original icing image from the front view. The proportion of bright pixels in the original icy image from a side view ; If the following conditions are met: Greater than or equal to the first threshold and Less than or equal to the second threshold, and, Greater than or equal to the first threshold and If the value is less than or equal to the second threshold, the pixel percentage of each ice type in the frontal view is output by segmenting the image from the frontal view, and the pixel percentage of each ice type in the side view is output by segmenting the image from the side view. If the following conditions are met: Greater than or equal to the first threshold and Less than or equal to the second threshold, and, Less than the first threshold or If the value is greater than the second threshold, the pixel percentage of each ice type in the frontal view segmentation image is output, and the pixel percentage of each ice type output in the frontal view is used for the side view. If the following conditions are met: Less than the first threshold or Greater than the second threshold, and, Greater than or equal to the first threshold and If the value is less than or equal to the second threshold, the pixel percentage of each ice type in the side view is output for image segmentation, and the pixel percentage of each ice type output from the side view is used for the front view. In other cases, the output will show that the recognition conditions are not met.

[0020] Preferably, the effective pixels include pixels of all icing types and pixels without icing.

[0021] Preferably, the icing density is calculated based on the pixel proportions of each icing type from the front view and the pixel proportions of each icing type from the side view; the icing thickness from the front view is calculated based on the segmented image from the front view; and the icing thickness from the side view is calculated based on the segmented image from the side view. The method is as follows: Based on the pixel proportion of each ice type in the frontal view segmentation image and the side view segmentation image, the density of each ice type is weighted and averaged to obtain the ice density. The number of effective pixels in the front view is obtained by segmenting the image from the front view; the ratio of the number of effective pixels in the front view to the number of pixels of the wire in the front view under ice-free conditions is multiplied by the wire diameter to obtain the ice thickness in the front view. The effective pixel count of the side view is obtained by segmenting the image based on the side view; the ice thickness of the side view is obtained by multiplying the ratio of the effective pixel count of the side view to the pixel count of the side view conductor in the ice-free state by the conductor diameter.

[0022] The present invention also provides a dual-view image icing monitoring device for transmission lines, comprising: A power module is disposed on the surface and inside of the device on the higher side of the transmission line conductor, parallel to the power supply line, and is used to provide power. The image acquisition module is set on the surface of the device perpendicular to the lower side of the transmission line conductor, and includes two cameras, one of which is set directly above the conductor and the other of which is set to the side of the conductor; the two cameras are used to simultaneously acquire the original icing images of the transmission line conductor from the front view and the original icing images from the side view. The meteorological data acquisition module is located inside the device on the lower side parallel to the transmission line conductor, and is used to collect meteorological information about the environment where the transmission line conductor is located. The main control module, located inside the device, is used to calculate the standard icing thickness using the aforementioned dual-view image icing monitoring method for transmission lines. The remote communication module is located inside the device on the upper side parallel to the transmission line conductor, and is used to send icing images, meteorological information and standard icing thickness to the monitoring master station.

[0023] Preferably, the power module includes: The solar panel is installed on the upper half of the device, parallel to the power transmission line conductor; Current transformers are installed inside the device, through which transmission line conductors pass; The battery is located inside the device.

[0024] Preferably, the energy-harvesting core of the current transformer is made of nanocrystalline material.

[0025] Preferably, the angle between the two cameras of the image acquisition module and the line connecting them to the center of the device surface is 90°.

[0026] Preferably, the image acquisition module further includes: a de-icing mechanism and a supplementary light disposed around the camera.

[0027] Preferably, the de-icing mechanism includes: a lens heating element, a mechanical ice-pushing mechanism, and a windshield wiper mechanism.

[0028] Compared with the prior art, the above-described technical solution of the present invention has the following advantages: This invention discloses a dual-view image icing monitoring method for power transmission lines. First, it acquires icing images from both the front and side views simultaneously captured by binocular cameras. Grayscale conversion and histogram equalization are used to dynamically enhance the feature contrast of icing images in different weather scenarios, resolving the problem of blurred icing details in complex environments. Second, high-level semantic features from both views, including fog and illumination features, are processed through channel attention and spatial attention mechanisms to generate deeply fused shared coding features. These shared coding features are then combined to generate segmented images from both views, effectively improving the accuracy of icing type identification. Next, the proportion of effective pixels and the proportion of bright pixels in the image are used as criteria to determine the quality of the icing image. The segmentation results of high-quality icing images are then used to optimize the dual-view segmentation results. This collaborative dual-view determination improves the accuracy of icing type identification in complex scenarios such as heavy fog and reflective exposure. Finally, the invention calculates icing density using the dual-view segmentation results and calculates the front-view icing thickness and the side-view icing thickness using a pixel ratio method, achieving the calculation of standard icing thickness. This invention, based on dual-view icing images, not only acquires more icing details in complex weather scenarios through complementary information from multiple perspectives, effectively improving the accuracy of icing type identification and thus enhancing the accuracy of icing density calculation; but also utilizes a conductor standard ice thickness extraction technique based on dual-view images to obtain icing thickness in two mutually perpendicular directions. By approximating the icing of irregularly shaped conductors as ellipses, the accuracy of icing cross-sectional area calculation is effectively improved. Furthermore, by combining the equal weight method to convert it into a circular cross-sectional standard icing, the accuracy of standard icing thickness calculation is ultimately enhanced. Attached Figure Description

[0029] To make the content of this invention easier to understand, the invention will be further described in detail below with reference to specific embodiments and accompanying drawings, wherein: Figure 1 A schematic diagram illustrating the principle of the standard icing thickness identification and calculation method; Figure 2 This is a schematic diagram for identifying ice thickness, where Figure 2 Image (a) in the image is of the conductor when there is no ice. Figure 2 (b) in the figure is the image of the conductor when there is ice; Figure 3 This is a schematic diagram illustrating the icing recognition effect in a single-view image. Figure 3 (a) in the image is the original icing image from a frontal view. Figure 3 (b) in the image is the original icing image from a side view. Figure 3 (c) in the image represents the recognition result for the frontal view alone. Figure 3 (d) in the figure represents the recognition result of the side view alone; Figure 4This is a flowchart of a dual-view image icing monitoring method for transmission lines according to the present invention; Figure 5 This is a comparison chart before and after histogram equalization. Figure 5 Image (a) is the grayscale icing image before histogram equalization. Figure 5 (b) in the image is the preprocessed image of the ice accretion after histogram equalization; Figure 6 A flowchart for the collaborative fusion of features from dual-view icing images; Figure 7 A structural diagram of the collaborative fusion of features from dual-view icing images; Figure 8 This is a structural diagram of a semantic segmentation network; Figure 9 A flowchart of one embodiment of the method of the present invention; Figure 10 This is a schematic diagram of the dual-view image icing monitoring device for transmission lines of the present invention installed on the conductor of a transmission line; Figure 11 This is a schematic diagram of the structure of the dual-view image icing monitoring device for power transmission lines of the present invention; Figure 12 A schematic diagram of the mechanical ice-pushing mechanism of a dual-view image icing monitoring device for power transmission lines during operation; Explanation of reference numerals in the accompanying drawings: 1. Power supply module; 2. Image acquisition module; 3. Weather acquisition module; 4. Main control module; 5. Remote communication module; 6. Front-view camera; 7. Side-view camera; 8. Mechanical ice-pushing mechanism; 9. Wiper mechanism; 10. Supplemental lighting; 11. Solar panel; 12. Power transmission line conductor; 13. Upper housing of the device; 14. Lower housing of the device. Detailed Implementation

[0030] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, so that those skilled in the art can better understand and implement the present invention. However, the embodiments described are not intended to limit the present invention.

[0031] Reference Figure 4 As shown, this invention provides a dual-view image icing monitoring method for transmission lines, the steps of which include: S1: Acquire the original icing images of the transmission line conductors from the front and side views, which were acquired synchronously. S2: Perform grayscale conversion and histogram equalization on the original icing images from the front and side views to obtain the preprocessed icing images from the front and side views. S3: Extract features from the preprocessed icing images from the front view and the preprocessed icing images from the side view, respectively, to obtain the high-level semantic features from the front view and the high-level semantic features from the side view; concatenate the high-level semantic features from the front view and the high-level semantic features from the side view and then pass them through the channel attention mechanism and the spatial attention mechanism in sequence to obtain the shared coding features. S4: Input the shared coding features, the high-level semantic features of the front view, and the preprocessed icing image of the front view into the semantic segmentation network to obtain the front view segmentation image; input the shared coding features, the high-level semantic features of the side view, and the preprocessed icing image of the side view into the semantic segmentation network to obtain the side view segmentation image. S5: Using the proportion of effective pixels in the segmented image and the proportion of bright pixels in the original icing image as the judgment indicators, the segmentation results of the front view segmented image and the side view segmented image are filtered and optimized to obtain the pixel proportion of each icing type in the front view and the pixel proportion of each icing type in the side view. S6: Calculate the icing density based on the pixel percentage of each icing type from the front view and the pixel percentage of each icing type from the side view; calculate the icing thickness from the front view based on the segmented image from the front view; calculate the icing thickness from the side view based on the segmented image from the side view; calculate the standard icing thickness based on the icing density, the icing thickness from the front view, and the icing thickness from the side view.

[0032] In S1, this embodiment uses a fixed dual-camera icing monitoring device. The synchronous triggering mechanism ensures the consistency of the shooting sequence and simultaneously collects images of the transmission line conductors at different angles from both the front and side views.

[0033] Specifically, in S2, grayscale conversion and histogram equalization are performed on the original icing images from the front and side views to obtain the preprocessed icing images from the front and side views. The steps include:

[0034] S21: Perform grayscale conversion on the original icing images from the front and side views to obtain grayscale icing images from the front and side views.

[0035] This embodiment performs grayscale conversion simultaneously on the original images from the front and side views, preserving key features of icing. By unifying the feature representation dimensions of the front and side views through grayscale conversion, redundant information in the color channels is removed, while accurately preserving the key features of three types of targets: rime, frost, and no icing. This establishes a unified visual benchmark for subsequent collaborative enhancement of foggy scenes and correct identification of icing types.

[0036] In this embodiment, when performing grayscale conversion, weights suitable for human visual observation, such as (R: 0.299, G: 0.587, B: 0.114), are used to perform a weighted average on the three channels to obtain a single-channel grayscale image. The specific formula is as follows: ; in, These are the coordinates of the pixels in the original icing image. For pixels grayscale value, , and Each pixel The RGB values.

[0037] S22: Obtain the icing area of ​​the conductor in the frontal view grayscale icing image and the side view grayscale icing image respectively, and divide them into several non-overlapping sub-blocks.

[0038] Specifically, in this embodiment, the conductor and its adjacent area are defined as the conductor icing area, and the pixels of the conductor icing area are denoted as... The blocks are divided into 8x8 pixel sections, with each 8x8 sub-block being non-overlapping. If some sub-blocks are smaller than 8x8 pixels, they are padded with mirror images to ensure that each sub-block reaches an 8x8 size.

[0039] S23: Calculate the grayscale histogram of each sub-block; select a preset limiting threshold based on the current weather scene, crop the pixels in each sub-block that exceed the limiting threshold, and count the number of cropped pixels in each sub-block; distribute the number of cropped pixels in each sub-block equally among all grayscale levels of the sub-block to obtain the enhanced grayscale histogram of the sub-block.

[0040] For each 8×8 sub-block, calculate the grayscale histogram of that sub-block. Where g∈[0,255] represents the gray level, and i and j are the row and column indices of the sub-block, respectively. The gray values ​​of each sub-block are counted to generate the corresponding histogram.

[0041] This embodiment presets corresponding climate values ​​for various climate scenarios (such as reflective exposure, heavy fog, or normal lighting). And based on each climate value, calculate the corresponding amplitude limiting threshold for each climate scenario. Each icing image selects a limiting threshold based on the current climate scene, and this limiting threshold is applied to each sub-block of the current icing image.

[0042] For the histogram of each sub-block If the pixel value corresponding to the grayscale value g exceeds the amplitude limit threshold If the pixel is not found to be cropped, then a cropping operation is performed on that pixel. After cropping, the number of cropped pixels in each sub-block is counted. .

[0043] The number of cropped pixels for each sub-block The gray levels of the sub-block are evenly distributed across all gray levels to obtain the enhanced gray-level histogram of that sub-block. This enhanced gray-level histogram is represented as follows: ; in, To enhance the grayscale histogram.

[0044] The goal of this process is to adjust the number of pixels at each gray level to make the gray level distribution of the image more balanced.

[0045] S24: Based on the enhanced grayscale histogram of each sub-block, map each pixel value of the sub-block to a new grayscale value to obtain the icing image after preprocessing from the front view and the icing image after preprocessing from the side view.

[0046] Specifically, calculate the enhanced grayscale histogram. cumulative distribution function Then through the cumulative distribution function Perform grayscale mapping to make the pixel value of each pixel in the icy area of ​​the conductor in the original grayscale icing image... Will be mapped to new grayscale values : ; in, This is for rounding operations.

[0047] The purpose of this step is to enhance the contrast of the image, making previously darker or brighter areas more uniform. See the image before and after histogram equalization for reference. Figure 5 As shown, where Figure 5 Image (a) is the grayscale icing image before histogram equalization. Figure 5 (b) in the image is the preprocessed image of the iced area after histogram equalization.

[0048] All histogram-equalized sub-blocks are stitched together to form a complete icing region on the conductor, resulting in a preprocessed icing image.

[0049] In S3, feature extraction is performed on the preprocessed icing images from the front view and the preprocessed icing images from the side view, respectively, to obtain the corresponding high-level semantic features from the front view and the side view. The method is as follows: The preprocessed icing image from the frontal view is input into the first feature extractor to obtain the frontal view high-level semantic features; the frontal view high-level semantic features include frontal view fog features. and positive angle lighting characteristics ; The preprocessed icing image from the side view is input into the second feature extractor to obtain high-level semantic features from the side view; the high-level semantic features from the side view include side view fog features. and side view lighting characteristics .

[0050] Preferably, the first feature extractor and the second feature extractor are DeepLabv3+, ResNet50 or MobileNetV3.

[0051] Specifically, refer to Figure 6 and Figure 7 As shown in S3, the high-level semantic features from the frontal view and the high-level semantic features from the side view are concatenated and then sequentially passed through a channel attention mechanism and a spatial attention mechanism to obtain shared encoded features. The steps include:

[0052] S31: Concatenate the high-level semantic features from the front view and the high-level semantic features from the side view to obtain the concatenated feature tensor.

[0053] In this embodiment, a unified target dimension is first set for the feature extractor: the number of channels per feature is C, and the spatial size is H×W. Then, convolutional layer parameters are designed for the original dimension of each feature, setting the number of input channels to the original number of input features, and fixing the output feature to the target number of channels C, thus achieving channel unification. Within the feature extractor, the kernel size, stride, and padding are adjusted simultaneously to extract four features, namely, the frontal view foggy feature. Orthogonal viewing angle lighting characteristics Side view characteristics of fog and side view lighting characteristics All dimensions have been adjusted to be consistent.

[0054] The splicing operation unfolds along the channel dimension of the features. The dimension of each single feature is C×H×W (number of channels × height × width). The four features are stacked sequentially along the channel dimension. The number of channels is accumulated from a single feature to 4C, while the spatial dimension H×W remains unchanged. The final feature splicing feature tensor has a complete dimension of 4C×H×W.

[0055] S32: The spliced ​​feature tensor is processed through a channel attention mechanism to obtain channel attention features. The steps include: The concatenated feature tensor is passed through a first convolutional layer, a second convolutional layer, a global average pooling layer, a first fully connected layer, a ReLU activation function, a second fully connected layer, and a Sigmoid activation function in sequence to obtain the channel attention weights; The channel attention features are obtained by performing a channel-wise element-wise multiplication operation on the channel attention weights and the concatenated feature tensor.

[0056] Convolutional layers use convolutional kernels to perform weighted summation of all channels at each spatial location of the concatenated feature tensor, outputting features with the same spatial size as the concatenated feature tensor. Global average pooling layers, on the other hand, calculate the average value of all pixels across the entire H×W spatial dimension for each channel of the input feature, compressing the spatial dimension of each channel from H×W to 1×1.

[0057] The first fully connected layer compresses the 4C channels to C using a preset compression ratio, and then introduces a non-linear transformation through the ReLU activation function. The activated features are then input into the second fully connected layer to restore the number of channels to 4C. Finally, the Sigmoid activation function normalizes the feature values ​​to the 0-1 range, resulting in channel attention weights with dimensions 4C×H×W. .

[0058] Channel attention weights Perform channel-wise element-wise multiplication with the concatenated feature tensor. The 0-to-1 weight value corresponding to each channel is multiplied one by one with the pixel value of all H×W spatial locations in the same channel in the concatenated feature tensor to obtain the channel attention feature. .

[0059] S33: The channel attention features are processed through a spatial attention mechanism to obtain shared features. The steps include: The channel attention features are concatenated after passing them through a global average pooling layer and a global max pooling layer, respectively, to obtain the fused features. The fused features are passed through a third convolutional layer and a sigmoid activation function in sequence to obtain spatial attention weights; The shared features are obtained by multiplying the spatial attention weights and the pixel values ​​corresponding to the channel attention features at their spatial locations one by one.

[0060] Specifically, channel attention features Simultaneously, global average pooling and global max pooling are performed. Global average pooling calculates the average pixel value across all 4C channels for each spatial location, compressing the channel dimension to 1 to obtain a 1×H×W feature map. Global max pooling extracts the maximum pixel value across all channels for each spatial location, similarly compressing the channel dimension to 1 to obtain a 1×H×W feature map. The 1×H×W feature map obtained from global average pooling and the 1×H×W feature map obtained from global max pooling are then concatenated along the channel dimension to form a fused feature map with a dimension of 2×H×W.

[0061] The fused features are passed through a 1×1 third convolutional layer and a sigmoid activation function to obtain spatial attention weights. Spatial attention weights The 0-to-1 weight value for each spatial location, and the channel attention feature. The pixel values ​​at this spatial location are multiplied one by one to generate a shared feature with a "channel + spatial weighting" dimension of 4C×H×W. .

[0062] S34: The shared coded features are obtained by adding the shared features and the concatenated features tensors and passing them through a convolutional layer.

[0063] Specifically, shared features The concatenated feature tensor is added to the concatenated feature tensor, and then subjected to final feature compression and enhancement through convolutional layers. The number of convolutional kernels is systematically adjusted to balance the richness of feature extraction and computational efficiency. At the same time, the kernel size is optimized: a 3×3 small kernel is used to enhance texture capture for local detail features, and a 5×5 large kernel is used to expand the receptive field for global structural features. Finally, the shared encoded features after co-coding are obtained. .

[0064] Shared coding features It deeply integrates multi-source information on fog and lighting from both front and side views, and is optimized with dual attention weighting, thus possessing accurate target discrimination capabilities. It can then be directly used as input to a semantic segmentation model for pixel-level prediction of icing-related data.

[0065] Preferably, in S4, the semantic segmentation network uses DeepLabv3+. DeepLabv3+ includes an encoder, an Atrous Spatial Pyramid Pooling (ASPP) module, and a decoder.

[0066] Specifically, refer to Figure 8 As shown, the shared coding features, frontal view high-level semantic features, and the frontal view preprocessed icing image are input into the semantic segmentation network to obtain the frontal view segmented image. The steps include:

[0067] The preprocessed icy image from the frontal view is input into the encoder of the semantic segmentation network to obtain the frontal view encoded features. The frontal view encoded features are then input into the hollow spatial pyramid pooling module to obtain the frontal view intermediate features. The shared encoded features are input into the hollow spatial pyramid pooling module to obtain the shared intermediate features. The frontal view intermediate features and the shared intermediate features are concatenated to obtain the frontal view context features. The frontal view context features are then concatenated with the frontal view high-level semantic features and passed sequentially through a Dropout layer, a normalization layer, and a ReLU activation function before being input into the decoder of the semantic segmentation network to obtain the frontal view segmented image.

[0068] Similar to the frontal view processing, the shared encoded features, side-view high-level semantic features, and the preprocessed side-view icing image are input into the semantic segmentation network to obtain the side-view segmented image. The steps include:

[0069] The preprocessed icy image from the side view is input into the encoder of the semantic segmentation network to obtain the side view encoded features. The side view encoded features are then input into the hollow spatial pyramid pooling module to obtain the side view intermediate features. The shared encoded features are input into the hollow spatial pyramid pooling module to obtain the shared intermediate features. The side view intermediate features and the shared intermediate features are concatenated to obtain the side view context features. The side view context features and the side view high-level semantic features are then concatenated and passed through a Dropout layer, a normalization layer, and a ReLU activation function, and then input into the decoder of the semantic segmentation network to obtain the side view segmented image.

[0070] Specifically, downsampling and channel expansion operations are performed in the encoder, downsampling the input feature map by 4x (channel expansion to 32), 8x (channel expansion to 96), and 16x (channel expansion to 320) to obtain encoded features. The encoded features and shared encoded features are then sequentially injected into the hollow spatial pyramid pooling module to capture the multi-scale context of the icing morphology.

[0071] The ASPP module in this embodiment contains 5 parallel branches, as follows: (1) 1×1 convolutional layer to capture fine edges; (2) 3×3 dilated convolution (rate=6) to perceive local icing features; (3) 3×3 dilated convolution (rate=12) to perceive local icing features in the middle layer; (4) 3×3 dilated convolution (rate=18) to perceive deep local icing features; (5) After image-level global average pooling, the image is passed through a 1×1 convolutional layer and bilinear upsampling is used to restore it to its original size, providing global information for the whole image.

[0072] The encoded features and shared encoded features are used as input features of the ASPP module. After passing through all branches, the output features of each branch are concatenated along the channel dimension to obtain the corresponding intermediate features and shared intermediate features. Then, the intermediate features and shared intermediate features are concatenated along the channel dimension to obtain the context features.

[0073] The contextual features and high-level semantic features output by the ASPP module are concatenated and integrated along the channel dimension. This cross-perspective feature aggregation operation yields the initial fused feature tensor. Subsequently, it passes through an overfitting suppression layer (Dropout), followed by normalization and a nonlinear transformation using the ReLU activation function, ultimately obtaining the input features for the decoder.

[0074] The input features of the decoder are fed into the decoder of the semantic segmentation network, and pixel-level segmentation is achieved through multi-stage processing: First, the feature map is deepened and non-linearly enhanced by two layers of 3×3 convolution, combined with batch normalization and activation function (ReLU); then, the feature map size is enlarged by 4 times through bilinear upsampling, and then the features are optimized and overfitting is suppressed by two layers of 3×3 convolution and overfitting suppression mechanism; finally, the feature channels are compressed to 4 dimensions through 1×1 convolution, and the probability of each category is calculated by Softmax activation function, and finally the pixel-level segmentation probability image is output, where the four channels correspond to the four segmentation targets: background, frost, rime, and no ice.

[0075] The dual-view segmentation image after icing contour and type recognition contains valid pixels and invalid pixels. Valid pixels include pixels of all icing types and pixels of uniced transmission line conductors. Invalid pixels refer to background pixels in the segmented image. In this embodiment, valid pixels refer to three types of target pixels: rime pixels, pixels of uniced transmission line conductors, and frost pixels, where rime pixels and frost pixels are icing type pixels.

[0076] Highlighted pixels refer to pixels whose grayscale / brightness values ​​in the original icing image exceed a set threshold (240).

[0077] In S5, the proportion of effective pixels in the segmented image and the proportion of bright pixels in the original icing image are used as judgment indicators to filter and optimize the segmentation results of the frontal view segmentation image and the side view segmentation image, obtaining the pixel proportions of each icing type from the frontal view and the pixel proportions of each icing type from the side view. The steps include: Calculate the percentage of effective pixels in a frontal view segmented image. The percentage of effective pixels in a side-view segmented image ; Calculate the proportion of bright pixels in the original icing image from the front view. The proportion of bright pixels in the original icy image from a side view ; If the following conditions are met: Greater than or equal to the first threshold and Less than or equal to the second threshold, and, Greater than or equal to the first threshold and If the value is less than or equal to the second threshold, the pixel percentage of each ice type in the frontal view is output by segmenting the image from the frontal view, and the pixel percentage of each ice type in the side view is output by segmenting the image from the side view. If the following conditions are met: Greater than or equal to the first threshold and Less than or equal to the second threshold, and, Less than the first threshold or If the value is greater than the second threshold, the pixel percentage of each ice type in the frontal view segmentation image is output, and the pixel percentage of each ice type output in the frontal view is used for the side view. If the following conditions are met: Less than the first threshold or Greater than the second threshold, and, Greater than or equal to the first threshold and If the value is less than or equal to the second threshold, the pixel percentage of each ice type in the side view is output for image segmentation, and the pixel percentage of each ice type output from the side view is used for the front view.

[0078] In other cases, the output will show that the recognition conditions are not met.

[0079] The target icing type is determined by the largest percentage of effective pixels in the segmented image. The effective pixels in the segmented image are then used as the outline of the icing guideline.

[0080] Specifically, the proportion of effective pixels in the segmented image is calculated using the following formula: ; in, The number of effective pixels in the image segmented by the viewpoint k. The number of pixels in the transmission line conductors when the viewpoint k is free of ice. From a positive perspective, This is a side view.

[0081] The proportion of highlighted pixels in the original icing image is calculated using the following formula: ; in, The number of highlighted pixels in the original icy image at viewpoint k.

[0082] Furthermore, for cases that "do not meet the identification criteria," further screening is performed: If the following conditions are met: Less than the first threshold and Less than the first threshold, and, Less than or equal to the second threshold or If the value is less than or equal to the second threshold, it is determined as "not meeting the recognition conditions - occluded"; If the following conditions are met: Greater than the second threshold and Greater than the second threshold, and, Greater than or equal to the first threshold or If the value is greater than or equal to the first threshold, it is determined as "not meeting the identification conditions - strong reflective exposure".

[0083] Preferably, the first threshold is 30% and the second threshold is 50%.

[0084] In traditional single-view icing type recognition mode, since only frontal view images can be acquired, the core rule for determining "not meeting the recognition conditions" is: if the ratio of effective pixels identified by the semantic segmentation algorithm to the total pixels of the conductor in the non-iced state is <30%, or the ratio of bright pixels on the conductor to the total pixels of the conductor is >50%. However, in dual-view icing type recognition, thanks to the supplementary side-view images, the judgment logic is significantly optimized: when the frontal view is affected by complex weather conditions such as overexposure, blurring, or shadow interference, making it impossible to determine the icing type, the side-view image will be automatically used for collaborative judgment; if the side-view image meets the criteria of "the ratio of effective pixels identified by semantic segmentation to the total pixels of the conductor in the non-iced state is ≥30%" and "the ratio of bright pixels on the conductor to the total pixels of the conductor is ≤50%", then the icing type recognition result from the side view is directly adopted, thereby improving the accuracy and effectiveness of icing type recognition.

[0085] On the other hand, in the traditional single-view mode, only the upper half of the conductor is visible, leading to a partial and inaccurate assessment of the actual pixel proportions of frost and ice, which can result in biased weighted averages of ice density. In contrast, the dual-view ice density calculation, through collaborative recognition of side-view images, accumulates the total pixels of frost and ice from both front and side views, calculates the pixel proportions of frost and ice, and then uses a weighted average to complete the ice density calculation, ensuring both the completeness and accuracy of the calculation.

[0086] In S6, the icing density is calculated based on the pixel proportions of each icing type from the front view and the pixel proportions of each icing type from the side view; the icing thickness from the front view is calculated based on the segmented image from the front view; and the icing thickness from the side view is calculated based on the segmented image from the side view. The method is as follows:

[0087] S61: Based on the pixel proportion of each ice type in the frontal view segmentation image and the side view segmentation image, the density of each ice type is weighted and averaged to obtain the ice density.

[0088] The complex climatic conditions in mountainous areas can lead to differences in image features between frontal and side views. For example, images taken from a frontal view at night may be overexposed due to the influence of supplementary lighting, while images taken from a side view may have normal image quality. To accurately determine the icing density, the dual-view collaborative method proposed in this invention simultaneously inputs two view images into a semantic segmentation model for refined type recognition. The model extracts texture, grayscale, and morphological features from each view. For cases where the icing process on the same conductor differs, such as a mixture of rime (density approximately 0.8 g / cm³) and hoarfrost (density approximately 0.2 g / cm³), the current icing density can be obtained by weighted averaging (hoarfrost pixel percentage × hoarfrost density + rime pixel percentage × rime density).

[0089] The percentage of frost pixels is calculated as follows: (total frost pixels from the front view + total frost pixels from the side view) / (total number of pixels for all icing types from the front view + total number of pixels for all icing types from the side view); the percentage of rime pixels is calculated as follows: (total rime pixels from the front view + total rime pixels from the side view) / (total number of pixels for all icing types from the front view + total number of pixels for all icing types from the side view). The current icing density is then obtained by weighted averaging.

[0090] For example, if the total number of pixels for all icing types from a frontal view is 1000 pixels, with rime occupying 400 pixels and glaze occupying 600 pixels; and the total number of pixels for all icing types from a side view is 500 pixels, with rime occupying 100 pixels and glaze occupying 400 pixels, then the final icing density = ((400+100) / (1000+500))*0.2 + ((600+400) / (1000+500))*0.8 = 0.599 g / cm³.

[0091] S62: Obtain the number of effective pixels from the front view based on the front view segmented image; multiply the ratio of the number of effective pixels from the front view to the number of pixels of the front view conductor in the ice-free state by the conductor diameter to obtain the front view icing thickness.

[0092] In the dual-view standard icing thickness calculation system, the frontal view icing thickness refers to the distance between the two farthest points on the ice surface of the conductor's iced cross-section as seen from the frontal view. First, a frontal view image of the conductor in an ice-free state is captured, and the number of pixels on the conductor in the ice-free state is recorded. Then, during the icing period, an iced image of the conductor is captured using the same lens and focal length. Based on the frontal view segmented image obtained from the above steps, the effective pixels in the segmented image are counted. The ratio of the number of effective pixels in the frontal view to the number of pixels on the conductor in the ice-free state is multiplied by the conductor diameter to obtain the frontal view icing thickness. .

[0093] If the number of effective pixels in the frontal view is greater than the number of pixels of the conductor in the frontal view under ice-free conditions, then the ice thickness in the frontal view is equal to the diameter of the conductor.

[0094] S63: Obtain the number of effective pixels in the side view based on the side view segmented image; multiply the ratio of the number of effective pixels in the side view to the number of pixels of the side view conductor in the ice-free state by the conductor diameter to obtain the side view ice thickness.

[0095] In a dual-view system, the side-view icing thickness refers to the distance between the two farthest points on the ice surface of the conductor's icy cross-section as seen in the side-view image after icing. Its calculation method is the same as the front-view icing thickness. A side-view image of the conductor in an icy state is captured, and the number of pixels of the conductor in the icy state is recorded. When icing occurs, an icy image of the conductor is captured using the same lens and focal length. Based on the side-view segmented image obtained from the above steps, the effective pixels in the side-view segmented image are counted. The ratio of the number of effective pixels in the side-view to the number of pixels of the conductor in the icy state is multiplied by the conductor diameter to obtain the side-view icing thickness. .

[0096] If the number of effective pixels in the side view is less than the number of pixels in the side view conductor in the ice-free state, then the ice thickness in the side view is equal to the conductor diameter.

[0097] Considering the irregularity of the actual icing morphology of transmission line conductors, the profile of the icing cross-section is first extracted and approximated as an elliptical icing cross-section that closely matches the actual profile for quantitative calculation. Then, based on the equal weight method, the elliptical cross-section icing is further equivalent to a density of 0.9 g / cm³. 3 The icing thickness of a circular cross-section is obtained by measuring the difference between the radius of the circle and the radius of the bare wire, which is the uniform icing thickness, i.e., the standard icing thickness.

[0098] The above steps are used to obtain the icing thickness from the frontal view. The icing thickness from the side view is Ice density is It adapts to the different quality characteristics of different types of icing.

[0099] Let the radius of the equivalent standard circle be the standard icing thickness. Adding the radius r of the bare wire, the area of ​​the circle minus the area of ​​the bare wire should equal the corrected area of ​​the ice cover: ; in, The standard icing thickness that needs to be determined is... Where is the radius of the bare wire. The density of the ice layer. The thickness of the ice layer is shown from the positive angle. The glazing thickness is shown from a side view. The left side of the equation represents a density of 0.9 g / cm³. 3 The weight of ice covering a circular cross-section is given by the equation on the right, which represents the actual density per unit length. The weight of ice covering an elliptical cross-section is used to derive the following standard formula for calculating ice thickness: .

[0100] Figure 9This is a flowchart of an embodiment of the method of the present invention, wherein: In step 1, (a) is the original icing image from the front view, and (b) is the original icing image from the side view. In step 2, (c) is the preprocessed icing image from the front view, and (d) is the preprocessed icing image from the side view. Step 3 is the collaborative fusion of dual-view image features. The purpose is to obtain shared coding features of deep fusion of dual-view fog and illumination information, so as to provide efficient feature input for pixel-level semantic segmentation of icy types. Step 4 involves icing contour and type recognition based on a semantic segmentation network (using DeepLabv3+ as an example). The goal is to input dual-view images into DeepLabv3+ and output pixel-level segmentation probability maps for four categories: background, rime, frost, and no icing, thereby recognizing icing contours and types. Specifically, (e) and (f) represent the frontal and side-view segmented images, respectively. Step 5 involves identifying icing types from both perspectives. Even if the proportion of bright pixels in the frontal view exceeds the second threshold, thanks to the supplementary constraints from the side-view image, the icing type cannot be accurately identified from the frontal view; the side-view information can still assist in the determination. The calculated determination index is... , , , According to the judgment rule of S5, the pixel ratio of each ice type in the side view is output by segmenting the image from the side view. The pixel ratio of each ice type in the side view is used for the front view. Finally, the ice type of the conductor is determined to be 100% rime and 0% glaze.

[0101] Step 6 is to obtain the standard icing thickness. According to the weighted average of S6, the icing density = frost pixel ratio × frost density + rain frost pixel ratio × rain frost density, that is, the icing density = 100% * 0.2 + 0% * 0.8 = 0.2 g / cm³, so the current icing density is 0.2 g / cm³; the icing thickness from the front view is a = 55.12 mm and the icing thickness from the side view is b = 45.95 mm; based on the dual-view thickness data, according to the standard icing thickness calculation formula: (Where r is the bare wire radius, r = 14.5 mm) Substituting the parameters, the standard icing thickness is calculated. =2.94mm.

[0102] In summary, this embodiment uses dual-view image fusion recognition to ultimately determine that the icing type of the transmission line conductor at that moment is rime, and the standard icing thickness is 2.94 mm.

[0103] This invention also provides a dual-view image icing monitoring device for transmission lines. The device is installed entirely on the transmission line conductor and is approximately cylindrical in shape, comprising: A power module is disposed on the surface and inside of the device on the higher side of the transmission line conductor, parallel to the power supply line, and is used to provide power. The image acquisition module is set on the surface of the device perpendicular to the lower side of the transmission line conductor, and includes two cameras, one of which is set directly above the conductor and the other of which is set to the side of the conductor; the two cameras are used to simultaneously acquire the original icing images of the transmission line conductor from the front view and the original icing images from the side view. The meteorological data acquisition module is located inside the device on the lower side parallel to the transmission line conductor, and is used to collect meteorological information about the environment where the transmission line conductor is located. The main control module, located inside the device, is used to receive data from the image acquisition module and the meteorological acquisition module, and to calculate the standard icing thickness using the aforementioned dual-view image icing monitoring method for transmission lines. The remote communication module is located inside the device on the upper side parallel to the transmission line conductor, and is used to send icing images, meteorological information and standard icing thickness to the monitoring master station.

[0104] The icing process on transmission lines is often accompanied by rain, snow, and fog. Image monitoring devices installed on the towers are often affected by fog and snow, making it difficult to clearly see the icing on the conductors even when within 10 meters of them. (Refer to...) Figure 10 As shown, by installing the device on the conductor and taking close-up (<0.5m) images of the conductor ahead, clear images of the conductor covered in ice can be obtained.

[0105] Specifically, the power module includes: The solar panel is installed on the upper half of the device, parallel to the power transmission line conductor; A current transformer is installed inside the device on the higher side of the transmission line conductor, perpendicular to it, through which the transmission line conductor passes. The storage battery is located inside the device on the higher side of the transmission line conductor, perpendicular to it. It is charged by solar energy obtained from solar panels and electrical energy induced by current transformers from the transmission line conductor.

[0106] The low temperatures, rain, snow, and freezing weather accompanying icing on transmission lines can cause conventional solar power systems with batteries to lose their charging source and experience battery energy degradation. This can lead to icing monitoring devices going offline for 2-3 days due to low power, rendering them ineffective. Installing the device on the conductor, in conjunction with a current transformer, can fully utilize the AC current transmitted by the conductor itself. Through electromagnetic induction, the device can generate energy for its monitoring purposes, ensuring high power supply reliability unaffected by severe weather such as rain and snow. Furthermore, arranging solar panels on the upper part of the device as auxiliary power supply addresses the issue of inductor failure due to light loads or short-term power outages, further improving power supply reliability.

[0107] Preferably, the image acquisition module further includes: a de-icing mechanism and a supplementary light disposed around the camera.

[0108] The de-icing mechanism includes: a lens heating element, a mechanical ice-pushing mechanism, and a windshield wiper mechanism.

[0109] The image acquisition module uses two cameras to observe the ice accretion on the conductor from different angles, improving the accuracy of ice accretion feature and cross-sectional area recognition calculations, thereby optimizing the calculation results of standard ice accretion thickness.

[0110] Lens fogging and icing in low-temperature rain, snow, and freezing weather is a major problem faced by image monitoring devices. The lens heating element is used to raise the lens temperature, the mechanical ice-pushing mechanism is used to remove ice from the lens's outer casing, and the wiper mechanism is used to remove rainwater from the lens. By combining these three components, the camera lens's resistance to rain, snow, and heavy fog can be improved, ensuring clear images even when iced. Icing on the wires typically grows faster at night; supplementary lights are placed near the two cameras to provide additional illumination for the lenses at night, ensuring clear images captured at night when icy.

[0111] Preferably, the meteorological data acquisition module includes ambient temperature and humidity sensors, which are arranged inside the device near the lower surface of the conductor parallel to the power transmission line. Collecting ambient temperature and humidity data can help determine icing information. Since the device is installed on a live conductor, to avoid corona discharge, the sensors are all placed inside the device, which has a smooth, curved shape.

[0112] Preferably, the current transformer's energy extraction core is made of nanocrystalline material.

[0113] The circuits where the device is installed will inevitably experience low current (<20A) operation. In this current range, the magnetization characteristics of conventional silicon steel cores are nonlinear, resulting in a large proportion of excitation current and a significant decrease in effective current output. To ensure effective induction power extraction under low current conditions, a nanocrystalline iron core with high permeability and low light-load loss is selected.

[0114] Preferably, the angle between the two cameras of the image acquisition module and the line connecting them to the center of the device surface is 90°.

[0115] Preferably, the lens heating element is activated when the weather acquisition module detects an ambient temperature ≤0℃ and humidity ≥80%.

[0116] The basic meteorological conditions for icing are a temperature ≤ 0℃, humidity ≥ 80%, and a certain wind speed. By controlling the activation of the lens heating element when the basic temperature and humidity conditions for icing are met, effective heating and de-icing can be ensured while saving energy.

[0117] Preferably, the mechanical ice-pushing mechanism is arranged in a circular shape around the two lenses. (Refer to...) Figure 11 andFigure 12 As shown, the mechanical ice-pushing mechanism is inside the device when it is not in operation, and moves outward along a direction parallel to the guide wire when it is in operation to push the ice.

[0118] The lens heating element primarily removes ice from the lens surface. However, when the device is in a windward position, the outer casing can quickly form a ring-shaped ice layer due to the continuous impact of supercooled water droplets, which gradually erodes inward and eventually completely obstructs the camera's field of view. Incorporating a mechanical ice-removing mechanism can effectively solve the problem of ice buildup on the outer casing obstructing the lens's view.

[0119] Preferably, the wiper mechanism is configured to correspond to the two cameras.

[0120] Besides the lens and the device housing being blocked by ice, the rain and snow falling on the lens during the icing process, or the water droplets formed after the heating element melts the ice, can also significantly affect the lens's shooting performance. This problem can be solved by adding a wiper mechanism.

[0121] To balance de-icing efficiency and energy utilization efficiency, and to achieve automatic de-icing activation to promptly remove ice that may affect lens shooting, this embodiment automatically controls the mechanical ice-pushing mechanism and the wiper mechanism based on ice image recognition results.

[0122] Preferably, the operating conditions of the mechanical ice-pushing mechanism and the windshield wiper mechanism are as follows:

[0123] The main control module detected that the proportion of effective pixels in the segmented image was less than the first threshold. or The mechanical ice-pushing mechanism moves first, followed by the wiper mechanism. After one action, images from both perspectives are collected again for recognition. If the recognition result still shows that "the proportion of effective pixels in the segmented image is less than the first threshold", the mechanical ice-pushing mechanism and the wiper mechanism move again in sequence until the proportion of effective pixels in the segmented image is greater than the first threshold.

[0124] Preferably, the supplementary light is automatically turned on when the image is recognized as nighttime.

[0125] The device proposed in this invention has multiple de-icing methods that work together to ensure the image monitoring effect of the device in rainy, snowy, and icy weather. By setting up three de-icing mechanisms—lens heating, mechanical ice pushing, and windshield wiper de-icing—it solves problems such as lens icing, outer shell icing and lens water obstruction in rainy, snowy, and icy weather, ensuring clear capture of images of icy conductors and improving monitoring results.

[0126] This invention achieves automatic control of the de-icing mechanism based on image recognition, improving de-icing efficiency. It determines whether the lens is obstructed based on the effective pixel ratio of ice monitoring in dual-view images, automatically starting and stopping mechanical ice pushing and windshield wiper de-icing, balancing de-icing efficiency and energy utilization efficiency, and promptly removing ice that affects lens shooting.

[0127] In summary, the dual-view image icing monitoring method for transmission lines described in this invention first acquires icing images from both the front and side views simultaneously captured by binocular cameras. Grayscale conversion and histogram equalization are used to dynamically enhance the feature contrast of icing images in different weather scenarios, resolving the problem of blurred icing details in complex environments. Secondly, high-level semantic features from both views, including fog and illumination features, are processed through channel attention and spatial attention mechanisms to generate deeply fused shared coding features. These shared coding features are then combined to generate segmented images from both views, effectively improving the accuracy of icing type identification. Next, the proportion of effective pixels and the proportion of bright pixels in the image are used as criteria to determine the quality of the icing image. The segmentation results of high-quality icing images are then used to optimize the dual-view segmentation results. This collaborative dual-view determination improves the accuracy of icing type identification in complex scenarios such as heavy fog and reflective exposure. Finally, this invention calculates icing density using the dual-view segmentation results and calculates the front-view icing thickness and the side-view icing thickness using a pixel ratio method, respectively, to achieve the calculation of standard icing thickness. This invention, based on dual-view icing images, not only acquires more icing details in complex weather scenarios through complementary information from multiple perspectives, effectively improving the accuracy of icing type identification and thus enhancing the accuracy of icing density calculation; but also utilizes a conductor standard ice thickness extraction technique based on dual-view images to obtain icing thickness in two mutually perpendicular directions. By approximating the icing of irregularly shaped conductors as ellipses, the accuracy of icing cross-sectional area calculation is effectively improved. Furthermore, by combining the equal weight method to convert it into a circular cross-sectional standard icing, the accuracy of standard icing thickness calculation is ultimately enhanced.

[0128] Obviously, the above embodiments are merely illustrative examples for clear explanation and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations here. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A method for monitoring icing on transmission lines using dual-view images, characterized in that, The method comprises the following steps: obtaining a front-view original icing image and a side-view original icing image of a power transmission line conductor synchronously collected; performing gray scale conversion and histogram equalization on the front-view original icing image and the side-view original icing image to obtain a front-view preprocessed icing image and a side-view preprocessed icing image; extracting features from the front-view preprocessed icing image and the side-view preprocessed icing image respectively to obtain front-view high-level semantic features and side-view high-level semantic features; concatenating the front-view high-level semantic features and the side-view high-level semantic features and then sequentially passing them through a channel attention mechanism and a spatial attention mechanism to obtain shared encoding features; inputting the shared encoding features, the front-view high-level semantic features and the front-view preprocessed icing image into a semantic segmentation network to obtain a front-view segmented image; inputting the shared encoding features, the side-view high-level semantic features and the side-view preprocessed icing image into the semantic segmentation network to obtain a side-view segmented image; taking the proportion of effective pixels in the segmented image and the proportion of highlighted pixels in the original icing image as judgment indexes, screening and optimizing the segmentation results of the front-view segmented image and the side-view segmented image to obtain the pixel proportions of each icing type in the front-view and the pixel proportions of each icing type in the side-view; calculating icing density based on the pixel proportions of each icing type in the front-view and the pixel proportions of each icing type in the side-view; calculating front-view icing thickness based on the front-view segmented image; calculating side-view icing thickness based on the side-view segmented image; and calculating standard icing thickness based on the icing density, the front-view icing thickness and the side-view icing thickness.

2. The method according to claim 1, wherein The method for performing gray scale conversion and histogram equalization on the front-view original icing image and the side-view original icing image to obtain the front-view preprocessed icing image and the side-view preprocessed icing image comprises the following steps: performing gray scale conversion on the front-view original icing image and the side-view original icing image to obtain a front-view gray scale icing image and a side-view gray scale icing image; obtaining conductor icing regions of the front-view gray scale icing image and the side-view gray scale icing image respectively, and dividing the conductor icing regions into a plurality of non-overlapping sub-blocks; calculating a gray scale histogram of each sub-block; selecting a preset clipping threshold based on a current meteorological scene, clipping pixels in each sub-block that exceed the clipping threshold, and counting the number of clipped pixels in each sub-block; and dividing the number of clipped pixels in each sub-block equally into all gray scales of the sub-block to obtain an enhanced gray scale histogram of the sub-block; mapping each pixel value of each sub-block to a new gray scale value based on the enhanced gray scale histogram of the sub-block to obtain the front-view preprocessed icing image and the side-view preprocessed icing image.

3. The method of claim 1, wherein the method further comprises: The method for extracting features from the front-view preprocessed icing image and the side-view preprocessed icing image respectively to obtain front-view high-level semantic features and side-view high-level semantic features comprises the following steps: inputting the front-view preprocessed icing image into a first feature extractor to obtain the front-view high-level semantic features; the front-view high-level semantic features comprise front-view foggy day features and front-view illumination features; The ice-coated image after the side-view angle preprocessing is input into a second feature extractor to obtain a side-view angle high-level semantic feature. The side-view angle high-level semantic feature includes a side-view angle foggy day feature and a side-view angle light feature.

4. The method according to claim 3, wherein The first feature extractor and the second feature extractor adopt DeepLabv3+, ResNet50 or MobileNetV3.

5. The method of claim 1, wherein, The positive-view angle high-level semantic feature and the side-view angle high-level semantic feature are spliced and sequentially subjected to a channel attention mechanism and a spatial attention mechanism to obtain a shared encoding feature, and the steps include: The positive-view angle high-level semantic feature and the side-view angle high-level semantic feature are spliced to obtain a spliced feature tensor; The spliced feature tensor is subjected to the channel attention mechanism to obtain a channel attention feature; The channel attention feature is subjected to the spatial attention mechanism to obtain a shared feature; The shared feature and the spliced feature tensor are added and subjected to a convolution layer to obtain the shared encoding feature.

6. The method of claim 5, wherein the method further comprises: The spliced feature tensor is subjected to the channel attention mechanism to obtain a channel attention feature, and the steps include: The spliced feature tensor is subjected to a first convolution layer, a second convolution layer, a global average pooling layer, a first full connection layer, a ReLU activation function, a second full connection layer and a Sigmoid activation function connected in sequence to obtain a channel attention weight; The channel attention weight and the spliced feature tensor are subjected to a channel-by-channel element multiplication operation to obtain the channel attention feature.

7. The method of claim 6, wherein the method further comprises: The channel attention feature is subjected to the spatial attention mechanism to obtain a shared feature, and the steps include: The channel attention feature is respectively subjected to a global average pooling layer and a global maximum pooling layer and then spliced to obtain a fusion feature; The fusion feature is subjected to a third convolution layer and a Sigmoid activation function connected in sequence to obtain a spatial attention weight; The spatial attention weight and pixel values of corresponding spatial positions of the channel attention feature are multiplied one by one to obtain the shared feature.

8. The method of claim 1, wherein, The semantic segmentation network adopts DeepLabv3+.

9. The method of claim 1, wherein, The shared encoding feature, the positive-view angle high-level semantic feature and the ice-coated image after the positive-view angle preprocessing are input into the semantic segmentation network to obtain a positive-view angle segmentation image, and the steps include: The ice-coated image after the positive-view angle preprocessing is input into an encoder of the semantic segmentation network to obtain a positive-view angle encoding feature; the positive-view angle encoding feature is input into a dilated spatial pyramid pooling module to obtain a positive-view angle intermediate feature; the shared encoding feature is input into the dilated spatial pyramid pooling module to obtain a shared intermediate feature; the positive-view angle intermediate feature and the shared intermediate feature are spliced to obtain a positive-view angle context feature; the positive-view angle context feature and the positive-view angle high-level semantic feature are spliced and sequentially subjected to a Dropout layer, a normalization layer and a ReLU activation function, and then input into a decoder of the semantic segmentation network to obtain the positive-view angle segmentation image; The shared encoding feature, the side-view angle high-level semantic feature and the ice-coated image after the side-view angle preprocessing are input into the semantic segmentation network to obtain a side-view angle segmentation image, and the steps include: The side view angle pre-processed icing image is input into an encoder of a semantic segmentation network to obtain side view angle coding features; the side view angle coding features are input into a hollow space pyramid pooling module to obtain side view angle intermediate features; the shared coding features are input into the hollow space pyramid pooling module to obtain shared intermediate features; the side view angle intermediate features and the shared intermediate features are spliced to obtain side view angle context features; the side view angle context features and the side view angle high-level semantic features are spliced, and then sequentially pass through a Dropout layer, a normalization layer and a ReLU activation function, and are input into a decoder of the semantic segmentation network to obtain a side view angle segmentation image.

10. The method of claim 1, wherein, The proportion of effective pixels in the segmentation image and the proportion of highlighted pixels in the original icing image are taken as the judgment indexes to screen and optimize the segmentation results of the front view angle segmentation image and the side view angle segmentation image, to obtain the pixel proportions of each icing type in the front view angle and the pixel proportions of each icing type in the side view angle, and the steps include: Calculate the proportion of effective pixels in the front-view angle segmented image , the proportion of effective pixels in the side-view angle segmented image ; calculate the proportion of highlight pixels in the front-view angle original icing image , the proportion of highlight pixels in the side-view angle original icing image ; If the following conditions are met: greater than or equal to a first threshold value and less than or equal to a second threshold value, and greater than or equal to a first threshold value and less than or equal to a second threshold value, then the image is divided into a front view angle and a side view angle, and the pixel proportion of each ice type in the front view angle and the pixel proportion of each ice type in the side view angle are output. If the following conditions are met: greater than or equal to a first threshold value and less than or equal to a second threshold value, and less than the first threshold value or greater than the second threshold value, the image is divided at the normal viewing angle to output the pixel proportion of each ice-coated type at the normal viewing angle, and the pixel proportion of each ice-coated type output at the normal viewing angle is used at the side viewing angle. If the following conditions are met: the first threshold value or the second threshold value, and the first threshold value and the second threshold value, the image is divided into side view angles, and the pixel proportion of each ice-covered type in the side view angle is output; the front view angle adopts the pixel proportion of each ice-covered type output by the side view angle. In other cases, the output is not satisfied with the recognition condition.

11. The method of claim 1, wherein, The effective pixels include the pixels of all icing types and the pixels of the conductor of the icing-free power transmission line.

12. The method of claim 1, wherein, The icing density is calculated based on the pixel proportions of each icing type in the front view angle and the pixel proportions of each icing type in the side view angle; The front view angle icing thickness is calculated based on the front view angle segmentation image; the side view angle icing thickness is calculated based on the side view angle segmentation image, and the method is: The density of each icing type is weighted and averaged based on the pixel proportions of each icing type in the front view angle segmentation image and the side view angle segmentation image, to obtain the icing density; The number of front view angle effective pixels is obtained based on the front view angle segmentation image; the ratio of the number of front view angle effective pixels to the number of pixels of the front view angle conductor in the ice-free state is multiplied by the conductor diameter to obtain the front view angle icing thickness; The number of side view angle effective pixels is obtained based on the side view angle segmentation image; the ratio of the number of side view angle effective pixels to the number of pixels of the side view angle conductor in the ice-free state is multiplied by the conductor diameter to obtain the side view angle icing thickness.

13. A transmission line dual-view image icing monitoring device, characterized by, It includes: A power module is arranged on the surface and inside of the device parallel to the higher side of the power transmission line conductor to provide power; An image acquisition module is arranged on the surface of the device perpendicular to the lower side of the power transmission line conductor, including two cameras, one of which is arranged directly above the conductor, and the other is arranged on the side of the conductor; The two cameras are used to synchronously acquire the front view angle original icing image and the side view angle original icing image of the power transmission line conductor; A weather acquisition module is arranged inside the device parallel to the lower side of the power transmission line conductor to acquire the weather information of the environment where the power transmission line conductor is located; A main control module is arranged inside the device to calculate the standard icing thickness by using the power transmission line double-view angle image icing monitoring method according to any one of claims 1 to 12; A remote communication module is arranged inside the device parallel to the upper side of the power transmission line conductor to send the icing image, weather information and standard icing thickness to the monitoring master station.

14. The dual-view image icing monitoring device for power transmission lines of claim 13, wherein, The power module includes: A solar panel is arranged on the upper half of the surface of the device parallel to the power transmission line conductor; A current transformer is arranged inside the device through which the power transmission line conductor passes; A storage battery is arranged inside the device.

15. The dual-view image icing monitoring device for power transmission lines of claim 14, wherein, The power taking core of the current transformer adopts nanocrystalline material.

16. The dual-view image icing monitoring device for power transmission lines of claim 13, wherein, The included angle between the two cameras of the image acquisition module and the line connecting the center of the device surface is 90 degrees.

17. The dual-view image icing monitoring device for power transmission lines of claim 13, wherein, The image acquisition module further comprises a deicing mechanism and a light supplement lamp arranged around the camera.

18. The dual-view image icing monitoring device for power transmission lines of claim 17, wherein, The deicing mechanism comprises a lens heating element, a mechanical ice pushing mechanism and a wiper mechanism.

Images