A method and system for detecting icing on a power transmission line

Abstract

The application provides a power transmission line icing detection method and system, and relates to the technical field of intelligent control. The method comprises the following steps: obtaining a clean transmission signal by eliminating transmission interference through a Manchester coding protocol and a local cache verification algorithm in a channel switching process after optimizing the signal; inputting the clean transmission signal into a pre-trained deep convolutional neural network, extracting ice layer spatial properties through multiple convolutional layers and pooling layers, performing feature fusion through a full connection layer, calculating ice layer thickness distribution, identifying the ice layer category through a softmax classifier, and generating a detection result containing ice layer thickness values and spatial distribution properties based on the thickness distribution map and the ice layer category identification result. The application can realize regional precise analysis and avoid local feature omission caused by traditional overall analysis.

Description

Technical Field

[0001] This invention relates to the field of intelligent control technology, and in particular to a method and system for detecting icing on power transmission lines. Background Technology

[0002] In areas with high altitude, low temperature and high humidity, power transmission lines are extremely vulnerable to icing. Icing disasters on power grids are sudden, especially during the rapid accumulation phase, when the ice layer can thicken by 2-3 mm per hour. Icing on lines not only increases the load on conductors and towers, causing excessive conductor sag, tower tilting or even collapse, but may also cause short circuits, tripping and other faults, seriously affecting the safe and stable operation of the power grid, and causing great disruption to social production and life.

[0003] Existing monitoring models have revealed significant shortcomings when dealing with such situations. For example, some models similar to ResNet-50, even with the support of high-performance servers, can only maintain a frame rate of 8-10 frames per second when processing 1080P resolution icing video (30 frames per second), resulting in a delay of 2-3 seconds. This delay prevents the monitoring system from tracking the ice thickening trend in real time. In real-world scenarios, by the time the monitoring system detects an icing anomaly, the actual icing situation may have already far exceeded expectations. Due to the inability to obtain accurate icing data in a timely manner, the dangerous condition of the transmission lines is underestimated, which may ultimately lead to the power towers becoming overloaded.

[0004] If drones are used for inspection, the latency issues of existing models will cause flight path planning to lag when drones transmit inspection results in real time. In the case of rapid icing, drones may be unable to avoid heavily iced tower areas in time, increasing the risk of collision with towers and failing to effectively complete the inspection task, further affecting the monitoring and response to power grid icing. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to provide a method and system for detecting icing on power transmission lines, which can achieve precise analysis by region and avoid the omission of local features caused by traditional overall analysis.

[0006] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows:

[0007] A method for detecting icing on power transmission lines, the method comprising:

[0008] Raw echo data from the surface of the transmission line is collected, and the raw echo data is preprocessed to obtain a preprocessed signal.

[0009] The preprocessed signal is converted into a set of discrete 3D data points. The Graham scan algorithm is used to extract the convex hull boundary of the set of discrete 3D data points to determine the peripheral morphological features of the data points. A midpoint closed curve generation algorithm is used to establish multiple reference ring templates within the peripheral morphological features. Based on the density distribution characteristics of the data points, the regions are divided to generate multiple feature analysis subdomains. Based on the curvature attributes and spatial arrangement of the data points in each feature analysis subdomain, principal component analysis is used to calculate spatial compensation coefficients. These spatial compensation coefficients are then used to correct the positional parameters of the data points to obtain an optimized processed signal.

[0010] The optimized signal is processed by using the Manchester coding protocol and local buffer verification algorithm during channel switching to eliminate transmission interference and obtain a clean transmission signal.

[0011] The clean transmission signal is input into a pre-trained deep convolutional neural network. The spatial attributes of the ice layer are extracted through multiple convolutional and pooling layers, and then the features are fused through a fully connected layer to calculate the ice layer thickness distribution. The ice layer category is identified by a softmax classifier. Based on the thickness distribution map and the ice layer category identification results, a detection result containing the ice layer thickness value and spatial distribution attributes is generated.

[0012] A power transmission line icing detection system, comprising:

[0013] The acquisition module is used to acquire raw echo data from the surface of the transmission line and to preprocess the raw echo data to obtain a preprocessed signal.

[0014] The correction module is used to convert the preprocessed signal into a set of discrete 3D data points, extract the convex hull boundary of the discrete 3D data point set based on the Graham scan algorithm, and determine the outer morphological features of the data points; establish multiple reference ring templates within the outer morphological features using a midpoint closed curve generation algorithm, divide the region according to the density distribution characteristics of the data points, and generate multiple feature analysis subdomains; calculate the spatial compensation coefficients using principal component analysis based on the curvature attributes and spatial arrangement of the data points in each feature analysis subdomain, and use the spatial compensation coefficients to correct the position parameters of the data points to obtain an optimized processed signal;

[0015] The elimination module is used to eliminate transmission interference and obtain a clean transmission signal during channel switching by using the Manchester coding protocol and local buffer verification algorithm.

[0016] The prediction module is used to input the clean transmission signal into a pre-trained deep convolutional neural network, extract the spatial attributes of the ice layer through multiple convolutional and pooling layers, perform feature fusion through a fully connected layer, calculate the ice layer thickness distribution, identify the ice layer category through a softmax classifier, and generate a detection result containing the ice layer thickness value and spatial distribution attributes based on the thickness distribution map and the ice layer category identification result.

[0017] The above-described solution of the present invention has at least the following beneficial effects:

[0018] By preprocessing the raw echo data from the transmission line surface, environmental noise (such as wind interference and electromagnetic clutter) in the echo signal can be effectively filtered out, invalid data points and abnormal fluctuations can be removed, and the data purity can be improved. Compared with the raw data without preprocessing, the preprocessed signal can reduce the sources of error in the subsequent feature extraction process, ensuring that the data can truly reflect the initial state of icing on the transmission line surface, providing reliable input for the subsequent conversion and analysis of three-dimensional discrete data point sets, and avoiding detection deviations caused by poor raw data quality.

[0019] Convex hull boundary extraction based on the Graham scanning algorithm can accurately capture the peripheral morphological features of a three-dimensional discrete data point set, clearly distinguish the transmission line body from the icing area, and avoid the problem of incomplete icing contour recognition. By generating a reference benchmark ring template through the midpoint closed curve and dividing the feature analysis subdomains by combining the data point density, it can achieve precise regional analysis for complex scenarios such as uneven icing thickness and local accumulation, avoiding the omission of local features caused by traditional overall analysis.

[0020] By using principal component analysis to calculate spatial compensation coefficients and correct errors in the location parameters of data points, the positional errors caused by factors such as ambient light and equipment angle deviations can be effectively offset, making the optimized signal more consistent with the real spatial properties of icing.

[0021] The Manchester coding protocol, with its built-in clock synchronization feature, can reduce signal aliasing during channel switching and mitigate the impact of electromagnetic interference and signal attenuation on data transmission. The local cache verification algorithm performs real-time verification of transmitted data, eliminating erroneous data and retransmitting lost information, ensuring a clean and distortion-free transmission signal. This optimized transmission scheme can meet the high timeliness requirements of data return for scenarios such as UAV inspection and remote real-time monitoring, avoiding the lag in detection results caused by transmission interference in existing methods. It provides stable data transmission guarantees for real-time UAV flight path planning and immediate icing risk warnings, reducing potential risks such as UAV collisions with towers and misjudgments of icing risks caused by transmission lag.

[0022] The efficient feature extraction of multiple convolutional and pooling layers and the fast feature fusion of fully connected layers significantly improve the data processing speed. Compared with the existing model's frame rate of 8-10 frames per second, it can increase the processing frame rate (such as supporting real-time processing of 1080P icing video at 30 frames per second), control the detection latency to the millisecond level, and realize real-time tracking of the rapid accumulation process of icing thickening by 2-3mm per hour, avoiding insufficient prediction of icing risk due to latency.

[0023] By using a softmax classifier to identify ice layer types (such as rime ice, glaze ice, and other ice types with different densities), multi-dimensional detection results are generated, including thickness values, spatial distribution, and ice layer type. Compared to existing methods that can only monitor thickness, this result can help staff more accurately determine the degree of hazard caused by ice (such as the greater destructive power of high-density glaze ice). Attached Figure Description

[0024] Figure 1 This is a schematic flowchart of a method for detecting icing on power transmission lines provided by an embodiment of the present invention.

[0025] Figure 2 This is a schematic diagram of a power transmission line icing detection system provided in an embodiment of the present invention. Detailed Implementation

[0026] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0027] like Figure 1 As shown, an embodiment of the present invention proposes a method for detecting icing on transmission lines, the method comprising the following steps:

[0028] Step S1: Collect raw echo data from the surface of the transmission line, and preprocess the raw echo data to obtain a preprocessed signal.

[0029] Step S2: The preprocessed signal is converted into a set of discrete 3D data points. The Graham scan algorithm is used to extract the convex hull boundary of the discrete 3D data points to determine the peripheral morphological features of the data points. Multiple reference ring templates are established within these peripheral morphological features using a midpoint closed curve generation algorithm. Regions are divided according to the density distribution characteristics of the data points to generate multiple feature analysis subdomains. Based on the curvature attributes and spatial arrangement of the data points within each feature analysis subdomain, a spatial compensation coefficient is calculated using principal component analysis. This spatial compensation coefficient is then used to correct the position parameters of the data points to obtain an optimized processing signal.

[0030] Step S3: For the optimized signal, during the channel switching process, the transmission interference is eliminated through the Manchester coding protocol and the local buffer verification algorithm to obtain a clean transmission signal.

[0031] Step S4: The clean transmission signal is input into a pre-trained deep convolutional neural network. The spatial attributes of the ice layer are extracted through multiple convolutional and pooling layers, and then the features are fused through a fully connected layer to calculate the ice layer thickness distribution. The ice layer category is identified through a softmax classifier. Based on the thickness distribution map and the ice layer category identification results, a detection result containing the ice layer thickness value and spatial distribution attributes is generated.

[0032] In this embodiment of the invention, the Graham scanning algorithm is used to accurately extract the convex hull boundary of the three-dimensional discrete data point set to determine the outer morphology. Combined with a midpoint closure curve generation algorithm, a feature analysis subdomain is divided. Then, principal component analysis is used to calculate spatial compensation coefficients to correct the data point position parameters, effectively reducing errors caused by the discreteness of the original data and spatial arrangement deviations. To enhance data transmission stability, a combination of Manchester coding protocol and local cache verification algorithm is used during channel switching. This actively eliminates transmission interference, ensuring that the optimized signal is transmitted to the backend analysis stage in a clean state, avoiding deviations in detection results due to transmission problems, and improving the reliability of data transmission. To achieve comprehensive icing analysis, a pre-trained deep convolutional neural network is used to deeply extract the spatial attributes of the ice layer through multiple convolutional and pooling layers. After feature fusion through a fully connected layer, it can accurately calculate the ice layer thickness distribution and accurately identify the ice layer category through a softmax classifier. This overcomes the limitation of traditional detection methods that can only measure thickness, providing more comprehensive information for icing hazard assessment. To improve the practicality of the test results, the final test results include both ice thickness values ​​and spatial distribution attributes. This not only visually displays the ice thickness of each section of the transmission line, but also clearly reflects the spatial distribution differences of the ice. This makes it easier for maintenance personnel to quickly locate high-risk icing sections, develop targeted de-icing and protection plans, and improve the efficiency of transmission line icing maintenance.

[0033] The specific implementation process of step S1 above can be divided into two core steps: raw echo data acquisition and preprocessing, as follows:

[0034] First, raw echo data is acquired. High-frequency radar sensors or ultrasonic echo detectors adapted to high-altitude and low-temperature environments are selected and installed at key monitoring locations on transmission line towers, covering high-risk areas for icing, such as sections with large conductor sag and the connection between towers and conductors. After starting the acquisition equipment, the acquisition frequency is set to match the rapid accumulation characteristics of icing, typically maintaining a frequency of 3-5 times per second to ensure real-time capture of the 2-3 mm ice thickening process per hour. During the acquisition process, current environmental parameters, including ambient temperature, air humidity, and wind speed, are recorded simultaneously to provide environmental reference for subsequent data correction. Finally, a set of raw echo signals containing both surface icing on the transmission line and characteristics of the conductor itself is obtained.

[0035] Next, preprocessing operations are performed on the raw echo data. The first step is noise suppression, using a smoothing filter to remove high-frequency noise generated by environmental electromagnetic interference and equipment fluctuations, retaining the effective echo components directly related to the line surface morphology (including icing). The second step is data screening, based on a preset threshold for the transmission line conductor outline range, removing invalid signals (such as reflections from trees and buildings) that exceed this range, reducing redundant data and improving subsequent processing efficiency. The third step is signal normalization, addressing the signal strength imbalance caused by distance differences between sensors at different acquisition locations, adjusting the intensity values ​​of all effective echo signals to a uniform range to ensure data consistency in subsequent conversions. Finally, signal integrity verification is performed to check for missing or abruptly changing segments in the processed signal. If missing data is found, it is supplemented using effective signals from adjacent acquisition periods, ultimately resulting in a preprocessed signal with concentrated effective information and stable quality.

[0036] In a preferred embodiment of the present invention, step S2 involves converting the preprocessed signal into a set of discrete 3D data points, extracting the convex hull boundary of the discrete 3D data points based on the Graham scan algorithm, and determining the peripheral morphological features of the data points, including:

[0037] Step S2-1: The preprocessed signal is converted into a set of discrete data points in three dimensions using a three-dimensional coordinate mapping function, wherein each data point contains spatial coordinate information;

[0038] Step S2-2: Sort the three-dimensional discrete data point set by polar angle, take the lowest point as the starting point, traverse all data points based on the Graham scan algorithm, determine the convexity of the point set by the vector cross product, and gradually construct the convex hull boundary.

[0039] Step S2-3: Based on the convex hull boundary extraction results, determine the peripheral morphological features of the data points, including the boundary point sequence and the convex hull polygon structure.

[0040] In this embodiment of the invention, the preprocessed signal is transformed into a set of discrete three-dimensional data points containing spatial coordinate information through a three-dimensional coordinate mapping function, allowing the originally abstract signal data to be presented in the form of concrete spatial points. The discrete three-dimensional data point set is first sorted by polar angle, and then, starting from the lowest point, all data points are traversed using the Graham scanning algorithm, and convexity is determined by vector cross product. This efficiently removes redundant data points, accurately constructs the convex hull boundary, avoids deviations in boundary extraction, and ensures the authenticity and reliability of the external morphological features. By extracting the convex hull boundary, the external morphological features containing the boundary point sequence and the convex hull polygon structure are determined. This not only determines the spatial distribution range of the data points, but also provides a precise spatial framework for the subsequent establishment of a reference benchmark ring template and the division of feature analysis subdomains, reducing errors caused by fuzzy morphological information in subsequent steps and ensuring the accuracy of the overall detection process.

[0041] In a preferred embodiment of the present invention, step S2-1 involves converting the preprocessed signal into a set of discrete three-dimensional data points using a three-dimensional coordinate mapping function, wherein each data point contains spatial coordinate information, including:

[0042] Step S2-1-1 involves performing time-domain analysis on the preprocessed signal to obtain the timestamp, amplitude, and frequency parameters of each sampling point in the signal. Specifically, this includes: determining the basic range of the time-domain analysis, using the complete data sequence of the preprocessed signal as the analysis object. This data sequence includes all valid echo signal segments after noise suppression, filtering, normalization, and integrity verification. Next, the time unit for time-domain analysis is set, using the sampling period during preprocessed signal acquisition as the smallest analysis unit. This sampling period matches the 3-5 times per second acquisition frequency set during the original echo data acquisition stage, ensuring that the analysis covers each independent signal sampling moment. Then, the timestamp of each sampling point is extracted, and the specific time information of each sampling point being acquired is recorded using the system clock of the associated acquisition device. The accuracy needs to be down to the millisecond level to accurately reflect the spatial changes of icing at different time points. Next, the amplitude parameters of each sampling point are acquired. Using time-domain waveform analysis tools, the peak signal strength at the corresponding timestamp for each sampling point is read. This peak value is the amplitude of the sampling point. The amplitude reflects the energy intensity of the echo signal reflected from the transmission line surface (including icing), indirectly demonstrating the reflection characteristics of the icing surface. Finally, the frequency parameters of each sampling point are calculated. The amplitude variation trend of multiple consecutive adjacent sampling points is analyzed. The signal fluctuation period is determined by identifying the time interval between two adjacent amplitude peaks. Based on the correspondence between period and frequency, the frequency of the signal segment where each sampling point is located is calculated, ensuring that the three core parameters—timestamp, amplitude, and frequency—are obtained for each sampling point.

[0043] Step S2-1-2: Based on the transformation relationship between the radar coordinate system and the spatial world coordinate system, the timestamp, amplitude, and frequency parameters of each sampling point are used as input. The X-axis, Y-axis, and Z-axis coordinates in three-dimensional space are calculated using a three-dimensional coordinate mapping function. Specifically, this includes: establishing the transformation relationship between the radar coordinate system and the spatial world coordinate system; using the tower position of the high-frequency radar sensor as a reference, determining the origin of the spatial world coordinate system (usually the center point of the sensor mounting bracket is set as the origin); and simultaneously determining the corresponding directions of each axis in both coordinate systems, where the radar coordinate system... The transmitting axis of the radar coordinate system corresponds to the direction parallel to the direction of the power transmission line conductor in the spatial world coordinate system (set as the X-axis). The horizontal receiving axis of the radar coordinate system corresponds to the horizontal direction perpendicular to the direction of the conductor in the spatial world coordinate system (set as the Y-axis). The vertical receiving axis of the radar coordinate system corresponds to the vertical direction in the spatial world coordinate system (set as the Z-axis). This correspondence determines the transformation rules between the two coordinate systems. Next, the input parameters for each sampling point are prepared. The timestamp, amplitude, and frequency parameters of each sampling point obtained in step S2-1-1 are organized into independent parameter groups, ensuring that the parameter groups for each sampling point are complete. The data is complete and without omissions. Then, the X-axis coordinate value of each sampling point is calculated. Combined with the echo signal propagation time reflected by the timestamp, and based on the speed of electromagnetic wave propagation in air, the straight-line distance from the radar sensor to the sampling point is calculated. Then, combined with the angle information in the X-axis direction (parallel to the direction of the conductor), this straight-line distance is decomposed into the X-axis direction to obtain the X-axis coordinate of the sampling point. When calculating the Y-axis coordinate value, similarly, based on the straight-line distance and the angle information in the Y-axis direction (the horizontal direction perpendicular to the direction of the conductor), the straight-line distance is decomposed into the Y-axis direction to obtain the Y-axis coordinate. When calculating the Z-axis coordinate value, the sensor's installation height is considered. The Z-axis coordinates are obtained by adjusting the installation height value, along with the angle decomposition results of the straight-line distance in the Z-axis direction (vertical direction). During the entire coordinate calculation process, the amplitude parameter needs to be used as an auxiliary reference. If the amplitude of a certain sampling point is significantly lower than the preset icing reflection intensity threshold, the angle information and propagation time calculation results of the sampling point need to be rechecked to ensure the accuracy of the coordinate calculation. At the same time, the frequency parameter is combined to determine whether the sampling points belong to the same icing area. If the frequency difference is too large, the application of the transformation relationship needs to be checked for deviations. Finally, it is ensured that each sampling point can obtain the corresponding X-axis, Y-axis and Z-axis coordinate values.

[0044] Step S2-1-3 aggregates the spatial coordinates calculated from all sampling points to generate a three-dimensional discrete data point set, where each data point is uniquely identified by its corresponding spatial coordinates. Specifically, this includes: performing a preliminary verification of the spatial coordinates of all sampling points, checking whether the X-axis, Y-axis, and Z-axis coordinate values ​​of each sampling point conform to the actual spatial range of the transmission line. For example, whether the X-axis coordinate is within the length range of the conductor in the current monitoring section, whether the Y-axis coordinate is within the reasonable monitoring area on both sides of the conductor, and whether the Z-axis coordinate is within the upper and lower fluctuation range of the actual erection height of the conductor. If the coordinate value of a sampling point is found to be outside the reasonable range, it is marked as an abnormal coordinate and the calculation process of step S2-1-2 is traced back again. After correction, the verification is performed again until the coordinate values ​​of all sampling points are within the reasonable spatial range. Next, the coordinate data is categorized and organized, grouped according to the transmission line location corresponding to the sampling points. For example, sampling points belonging to the same conductor or the same tower connection section are grouped together. Within each group, the sampling points are arranged in chronological order according to their timestamps, facilitating subsequent viewing of the spatial distribution of sampling points at different times and locations. Then, the basic structure of the three-dimensional discrete data point set is constructed. The X-axis, Y-axis, and Z-axis coordinate values ​​of each verified sampling point are combined into an independent spatial coordinate data unit, and each spatial coordinate data unit is a three-dimensional discrete data point. Finally, to ensure the uniqueness of the data points, a duplicate check is performed on all constructed spatial coordinate data units. If two data points are found to have identical X-axis, Y-axis, and Z-axis coordinate values, the reason for the duplication is analyzed. If it is due to calculation error, the coordinate value of one data point is corrected; if it is due to duplicate sampling, the data point with the earlier timestamp is retained. Finally, all unique spatial coordinate data units are integrated together to form a complete three-dimensional discrete data point set. Each data point can be mapped to a specific spatial location on the surface of the transmission line through its unique X-axis, Y-axis, and Z-axis coordinate values.

[0045] In this embodiment of the invention, step S2-1-1 performs time-domain analysis on the preprocessed signal to extract the timestamp, amplitude, and frequency parameters of each sampling point, transforming the originally abstract echo signal into quantifiable specific parameters. This provides a definite data basis for subsequent spatial coordinate transformation, making the signal characteristics related to icing of transmission lines easier to analyze and utilize. Step S2-1-2 calculates the three-dimensional coordinates based on the transformation relationship between the radar coordinate system and the spatial world coordinate system, combined with the timestamp, amplitude, and frequency parameters of the sampling points. This accurately corresponds the signal parameters to the actual spatial location of the transmission line, avoiding coordinate errors caused by coordinate system transformation deviations, and ensuring that the spatial location of each data point truly reflects the actual distribution of icing and the line.

[0046] To ensure the validity of the data point set, step S2-1-3 aggregates the spatial coordinates of all sampling points. Through unique identification and integration, a three-dimensional discrete data point set is formed. This not only eliminates duplicate or invalid coordinate data, but also ensures that the data point set can completely cover the monitoring area. This provides high-quality data support for subsequent operations such as Graham scanning algorithm to extract convex hull boundaries and divide feature analysis subdomains, reducing detection errors caused by data problems in subsequent steps.

[0047] In a preferred embodiment of the present invention, step S2-2 involves sorting the three-dimensional discrete data point set by polar angle, using the lowest point as the starting point, traversing all data points based on the Graham scan algorithm, determining the convexity of the point set through the vector cross product, and gradually constructing the convex hull boundary, including:

[0048] Step S2-2-1: Select the point with the smallest Z-coordinate value from the set of discrete 3D data points as the origin. If multiple points have the same minimum Z-coordinate value, select the point with the smallest Y-coordinate value. If multiple points still have the same minimum Y-coordinate value, select the point with the smallest X-coordinate value as the final starting point. Specifically, this includes: obtaining the complete coordinate information of all data points in the set of discrete 3D data points. Each data point contains coordinate values ​​in three directions: X-axis, Y-axis, and Z-axis. The three coordinate values ​​of each data point need to be read and recorded one by one to form a list containing the coordinates of all data points. Filter the data points with the smallest Z-coordinate value. Compare the Z-coordinate values ​​of all data points in the list one by one to find all data points with the smallest Z-coordinate value. If only one data point in the list has the smallest Z-coordinate value, directly use that data point as a candidate starting point. If multiple data points have the same minimum Z-coordinate value, perform a second comparison on the Y-coordinate values ​​of these data points to filter out the data point with the smallest Y-coordinate value. If only one data point has the smallest Y-coordinate value, use that data point as a candidate starting point. If, after filtering by Z and Y coordinates, multiple data points still have the same minimum Z and Y coordinate values, then the X coordinate values ​​of these data points are compared three times to select the data point with the smallest X coordinate value. This data point is the final determined starting point, ensuring that the starting point has a unique and definite position in three-dimensional space, providing a unified reference benchmark for subsequent polar angle sorting.

[0049] Step S2-2-2: Using the starting point as the pole, calculate the polar angle of all other data points relative to this pole, and sort the data points in ascending order of polar angle to form a sorted data point sequence. Specifically, this includes: determining the reference direction for polar angle calculation; using the starting point as the pole, setting the positive X-axis direction parallel to the transmission line conductor direction as the reference direction for polar angle calculation (i.e., the polar axis direction). This direction is consistent with the X-axis direction of the spatial world coordinate system in previous steps, ensuring that the polar angle calculation matches the actual spatial layout of the transmission line; then, calculating the polar angle of each non-starting point data point relative to the starting point. For each data point to be calculated, using the starting point as the vertex, construct a vector pointing from the starting point to that data point, and a vector pointing from the starting point to the data point. The point points to a vector in the polar direction (positive X-axis direction). By analyzing the angle between these two vectors in space, the polar angle of the data point relative to the starting point is determined. The polar angle must cover the range of 0 to 360 degrees to ensure that data points in different directions can be distinguished. Then, all non-starting point data points are sorted according to the size of their polar angles. The polar angles of each data point are compared one by one, and the data points are arranged in ascending order of polar angle. If two or more data points have the same polar angle, the straight-line distance from these data points to the starting point is compared, and the data points farther from the starting point are placed first to avoid missing outer data points when constructing the convex hull boundary due to the same polar angle. Finally, a sequence of data points arranged in order of polar angle is formed.

[0050] Step S2-2-3: Initialize an empty stack. Push the starting point and the first two data points in the sorted sequence onto the stack in sequence. Then, traverse the remaining data points in sorted order. For each current traversed point, calculate the cross product of the second point on the top of the stack, the stack vertex, and the current traversed point. Based on the cross product result, determine whether the current traversed point causes the stack vertex to be a non-convex hull vertex. If so, pop the stack vertex. Continue until the stack top satisfies the convexity condition, then push the current traversed point onto the stack. This gradually constructs the convex hull boundary. Specifically, this includes: initializing an empty stack to store candidate points of the convex hull boundary. This stack needs to have the function of storing and popping data points in sequence to ensure that the stack top and the second data point on the top of the stack can be easily obtained in subsequent operations. Next, push the previously determined starting point onto the stack. Then, take the first two data points from the sorted data point sequence and push them onto the stack in the order of the sequence. At this time, the stack has stored three data points, forming the initial basis of the convex hull boundary. Then, following the sorted sequence of data points, traverse all remaining data points starting from the third data point. For the data point currently being traversed (i.e., the current traversal point), first obtain the two data points at the top of the stack, with the one closer to the stack entrance being the stack vertex and the one below the stack vertex being the second point at the top of the stack. Next, two vectors are constructed: the first vector points from the second point on the stack to the top of the stack, and the second vector points from the top of the stack to the current traversal point. The cross product is calculated by analyzing the directional relationship between these two vectors in space. If the cross product result indicates that the current traversal point will cause the top of the stack to be in a non-convex hull position (i.e., the angle formed by the two vectors is not acute, causing the top of the stack to be concave inward, which does not meet the characteristic of the convex hull boundary being convex outward), then the top of the stack is popped from the stack. The cross product of the vector formed by the second point on the stack after popping, the new top of the stack, and the current traversal point is checked again. The popping operation is repeated until the vector formed by the two points on the stack and the current traversal point satisfies the convexity condition (i.e., the cross product result indicates that the angle is acute, and the top of the stack is in a convex hull boundary position). Finally, the current traversal point is pushed onto the stack. Through this point-by-point traversal and judgment, data points that meet the requirements of the convex hull boundary are gradually filtered and retained, thus constructing a preliminary convex hull boundary.

[0051] Step S2-2-4: After the traversal is complete, the point sequence stored in the stack is the final extracted convex hull boundary data points. Specifically, this includes: continuously processing all remaining data points in the sorted data point sequence according to the traversal method in step S2-2-3, ensuring that the stack stores candidate convex hull boundary points that meet the convexity condition at the current stage after processing each data point; when the traversal operation is complete, stop processing the data points. At this point, there are no more data points in the stack that need further judgment and popping, and the data points stored in the stack have formed a complete sequence arranged sequentially from the starting point along the convex hull boundary; then, process the data points stored in the stack... The stored data point sequence undergoes an integrity check to confirm whether it contains all the key data points protruding outward from the periphery of the three-dimensional discrete data point set. This ensures that no important boundary points reflecting the outer morphology of the transmission line icing are omitted, and that no redundant non-boundary points are included. Finally, the data point sequence stored in the stack is determined to be the final extracted convex hull boundary data points. When these data points are connected in sequence, they can accurately present the outer contour of the three-dimensional structure formed by the ice on the transmission line surface and the conductor body. This provides a precise boundary framework for subsequent establishment of a reference benchmark ring template and division of feature analysis subdomains, ensuring the accuracy of subsequent icing detection steps.

[0052] In this embodiment of the invention, step S2-2-1 uses the minimum value sequentially to select a unique starting point based on the Z-axis, Y-axis, and X-axis coordinates. This avoids deviations in subsequent polar angle sorting and convex hull construction due to an uncertain starting point, providing a unified and reliable reference benchmark for the entire convex hull boundary extraction and adapting to the three-dimensional distribution characteristics of transmission line icing data points. Step S2-2-2 calculates the polar angles using the starting point as the pole and sorts them according to the polar angles, ensuring that the data points are arranged in an orderly manner according to spatial directions. This avoids confusion during subsequent data point traversal, and the polar angle sorting can cover different directions in three-dimensional space. Ice-covered data points; Step S2-2-3 uses stack operations combined with vector cross product to determine convexity, and removes data points that are not convex hull vertices one by one, retaining only valid vertices that can form the outer convex contour. This effectively avoids boundary extraction errors caused by the discreteness and irregularity of data points on the ice-covered surface, ensuring that the convex hull boundary can truly reflect the outer shape of the ice and the line; Step S2-2-4 determines the final convex hull boundary data points by traversing the point sequence in the stack, ensuring that the boundary point sequence is continuous and covers the key positions on the periphery of the monitoring area, reducing deviations caused by inaccurate boundary information in subsequent detection steps.

[0053] In a preferred embodiment of the present invention, step S2-3, based on the convex hull boundary extraction result, determines the peripheral morphological features of the data points, including the boundary point sequence and the convex hull polygon structure, including:

[0054] Step S2-3-1: Read the convex hull boundary data point sequence and connect each data point sequentially according to its storage order in the stack to form an initial boundary polygon. Specifically, this includes: reading the convex hull boundary data point sequence stored in the stack, extracting the X-axis, Y-axis, and Z-axis coordinate information of each data point one by one according to the storage order in the stack, ensuring that the extraction order is completely consistent with the storage order in the stack, and that no data points are missed or reversed. Next, starting from the first data point in the sequence, connect the current data point to the next data point sequentially using spatial line segments. When connecting, ensure that the line segments strictly correspond to the spatial coordinates of the two data points and do not deviate from the actual position of the data points. When connecting to the last data point in the sequence, connect the last data point to the first data point in the sequence to form a closed polygon structure, which is the initial boundary polygon. During the connection process, the spatial direction and length of each connecting line segment need to be recorded simultaneously for subsequent verification of the shape of the initial boundary polygon, ensuring that the initial boundary polygon can completely cover the spatial area enclosed by the convex hull boundary data points and accurately reflect the preliminary shape of the outer contour of the transmission line icing and the conductor body.

[0055] Step S2-3-2 involves optimizing the initial boundary polygon by removing collinear redundant vertices and retaining key turning vertices to generate an optimized boundary point sequence. Specifically, this includes: traversing all vertices of the initial boundary polygon and selecting three consecutive vertices in sequence according to their order within the polygon, denoted as the preceding vertex, current vertex, and subsequent vertex; then analyzing the spatial relationship between these three vertices, identifying collinear vertices by determining whether the line segment from the preceding vertex to the current vertex and the line segment from the current vertex to the subsequent vertex lie on the same straight line; specifically, observing the coordinate change pattern of the three vertices, and determining if the ratio of the difference in X-axis coordinates between the current vertex and the preceding vertex to the difference in Y-axis coordinates is similar to that between the subsequent vertex and the current vertex. If the ratio of the difference between the X-axis coordinate and the difference between the Y-axis coordinate is exactly the same, and the change in the Z-axis coordinate also shows the same proportional relationship, then the three vertices are determined to be collinear, and the current vertex is a redundant vertex. Then, the identified collinear redundant vertices are removed from the vertex sequence of the initial boundary polygon, and only the preceding and following vertices, which are key vertices that can reflect the turning shape of the polygon, are retained. The above operation of selecting three consecutive vertices to determine collinearity and removing redundant vertices is repeated until all vertices of the initial boundary polygon have been traversed, and finally an optimized boundary point sequence is formed. After optimization, the integrity of the sequence needs to be checked again to ensure that no key turning vertices are missed, and that the connection of adjacent vertices can accurately present the convex shape of the polygon, so as to avoid the distortion of the boundary shape due to the removal of redundant vertices.

[0056] Step S2-3-3: Based on the optimized boundary point sequence, calculate its geometric center point, and use this center point as a reference to emit rays to each boundary point, connecting the boundary point sequence to form a closed convex hull polygon structure. Specifically, this includes: calculating the geometric center point of the optimized boundary point sequence; summing the X-axis coordinates of all boundary points in the sequence and dividing by the total number of boundary points to obtain the X-axis coordinate of the geometric center point; using the same method, summing the Y-axis and Z-axis coordinates of all boundary points and dividing by the total number of boundary points to obtain the Y-axis and Z-axis coordinates of the geometric center point, thereby determining the complete spatial coordinates of the geometric center point; then, using this geometric center point as a reference, radiating rays to each boundary point in the optimized boundary point sequence... Each ray is launched from a geometric center point to a corresponding boundary point, ensuring that each ray points precisely to its corresponding boundary point without deviation. Then, following the optimized boundary point sequence, adjacent boundary points are connected sequentially by line segments. The line segments must spatially match the previously launched rays to form a continuous boundary profile. Finally, the last boundary point in the sequence is connected to the first boundary point, creating a closed structure—the convex hull polygon structure. During construction, the closure and integrity of the convex hull polygon structure must be checked to ensure it accurately encloses the outer region of the discrete three-dimensional data point set, truly reflecting the spatial periphery of the transmission line icing.

[0057] Step S2-3-4 involves outputting the boundary point sequence and convex hull polygon structure as the peripheral morphological features of the data points. Specifically, this includes: organizing the optimized boundary point sequence by arranging the X, Y, and Z axis coordinates of each boundary point into structured data in sequence order, while recording the total number of boundary points and the line segment lengths and spatial angles between adjacent boundary points, forming a complete boundary point sequence information package to ensure rapid retrieval of the position and sequence association information of each boundary point during subsequent calls; next, recording the attributes of the convex hull polygon structure, including its spatial volume, surface area, and the shape and size of each face, while also annotating the geometry. The coordinates of the center point are used to determine the spatial parameters of the convex hull polygon structure, making its morphological features quantifiable. Then, the organized boundary point sequence information is integrated with the convex hull polygon structure attribute information to form a unified peripheral morphological feature dataset. During integration, it is crucial to ensure the spatial correspondence of the two types of information is consistent, preventing information misalignment. Finally, this peripheral morphological feature dataset is output to the subsequent processing module, specifically for establishing a reference ring template using the midpoint closed curve generation algorithm. During output, the integrity of the data transmission must be checked to ensure that the subsequent module can fully receive all information from the boundary point sequence and the convex hull polygon structure.

[0058] In a preferred embodiment of the present invention, step S2-3-1 connects the convex hull boundary data points in the stack order to form an initial boundary polygon, which can completely restore the outer contour of the three-dimensional discrete data point set, providing a basic framework for subsequent morphological optimization and structural construction, and ensuring that the initial appearance of the outer morphology of the icing and the line is complete. Step S2-3-2 removes collinear redundant vertices in the initial boundary polygon, reducing the amount of invalid data to improve the efficiency of subsequent processing, while retaining key turning vertices to avoid morphological distortion caused by redundant vertices, ensuring that the optimized boundary point sequence can accurately reflect the convex and turning features of the outer morphology. Step S2-3-3 constructs a closed convex hull polygon structure by calculating the geometric center point, emitting rays and connecting the boundary points, so that the outer morphology is transformed from a point sequence into a three-dimensional structure. Step S2-3-4 integrates the boundary point sequence and the convex hull polygon structure for output, forming standardized and complete outer morphological feature data, ensuring that subsequent steps (such as midpoint closed curve generation and spatial compensation coefficient calculation) can directly call accurate morphological information, reducing subsequent detection errors caused by non-standard morphological data, and ensuring the consistency and accuracy of the overall detection process.

[0059] In a preferred embodiment of the present invention, multiple reference ring templates are established within the peripheral morphological features using a midpoint closed curve generation algorithm. Based on the data point density distribution characteristics, the region is divided to generate multiple feature analysis subdomains, including:

[0060] Step S2-4: Based on the geometric center point of the convex hull polygon structure, emit probe rays radially at equal angles to the boundary of the convex hull, and obtain the coordinates of the intersection points of each ray and the boundary. Specifically, this includes: determining the coordinates of the geometric center point of the convex hull polygon structure, which has been calculated in step S2-3-3 and needs to be accurately read and used as the origin of the probe rays; then, setting the equal angle interval for emitting probe rays radially. Considering the circular or near-circular characteristics of the outer morphology of the icing periphery of the transmission line, the angle interval is usually set to 15 degrees to 30 degrees to ensure that the rays can uniformly cover the entire circumference of the convex hull polygon and avoid detection blind spots; then, starting from the geometric center point, emit probe rays sequentially according to the set equal angle intervals, each... Each ray extends radially outward until it intersects the boundary of the convex hull polygon. During the ray's extension, the spatial relationship between the ray trajectory and the convex hull boundary data points is compared in real time. When the ray trajectory passes through a convex hull boundary data point, the coordinates of that data point are recorded, and these coordinates are the intersection coordinates of the ray and the convex hull boundary. If the ray trajectory does not directly pass through a convex hull boundary data point, the two convex hull boundary data points closest to each other on both sides of the ray are selected, and the intersection coordinates of the ray and the convex hull boundary are calculated by analyzing the spatial relationship between the two points and the ray. The above operation of emitting rays and determining intersection coordinates is repeated until all rays with equal angular intervals have found their corresponding intersections, and finally, the set of intersection coordinates of all rays and the convex hull boundary is obtained.

[0061] Steps S2-5 involve generating multiple equally spaced segment points on each ray according to a preset ratio, using the geometric center point as a reference. Connecting segment points of the same ratio on adjacent rays forms a closed loop curve, constructing multiple concentrically distributed reference loop templates. Specifically, this includes: determining the preset ratio and the number of segment points; setting multiple preset ratio values, such as 5%, 10%, 20% up to 95%, based on the accuracy requirements for detecting the icing thickness of transmission lines, with each ratio corresponding to the radius percentage of a loop template; simultaneously determining the number of segment points generated on each ray based on the actual size of the convex hull polygon, ensuring equal spacing between adjacent segment points and covering the entire ray length from the geometric center point to the intersection of the convex hull boundary; then generating segment points on each ray, with the total length being the distance from the geometric center point to the intersection of the ray and the convex hull boundary. The process involves calculating the distance from each segment point to the geometric center point according to a preset ratio. Then, based on the coordinate variation pattern of the ray direction, the X, Y, and Z coordinates of each segment point are calculated, thus generating all segment points on a single ray. Next, segment points of the same ratio on adjacent rays are connected. Following the preset ratio from smallest to largest, segment points of the same ratio on all rays are selected sequentially and connected with line segments according to the angular order of the rays, forming a closed loop curve. This operation is repeated to generate a corresponding closed loop curve for each preset ratio, ultimately constructing multiple reference ring templates concentrically distributed with the geometric center point as the center. After generation, the closure and integrity of each ring template must be checked to ensure smooth connection between adjacent segment points and that the ring template accurately reflects the spatial range of different radial positions.

[0062] Steps S2-6 involve statistically analyzing the data point density distribution within each annular template, determining density abrupt change boundaries through density gradient analysis, and dividing the entire region into multiple feature analysis subdomains with similar density characteristics based on density distribution characteristics. Specifically, this includes: counting the number of data points within each annular template; traversing all data points in the three-dimensional discrete data point set; determining whether each data point is located within the region enclosed by the current annular template (i.e., the distance from the data point to the geometric center point is within the radius range of the current annular template and its inner adjacent annular template); if located within the region, counting it as part of the total number of data points within that annular template; and then calculating the data point density within each annular template, using the area enclosed by each annular template (…). Using the volume (or volume, depending on the three-dimensional space) as the denominator and the total number of data points within the region as the numerator, the density value of each region is calculated to form a density distribution sequence. Then, density gradient analysis is performed to calculate the density difference between two adjacent annular template regions. If the density difference between two adjacent regions exceeds a preset density threshold (this threshold is set according to the normal distribution characteristics of icing data points), the boundary between the adjacent regions is determined to be a density abrupt change boundary. Finally, feature analysis subdomains are divided according to the density abrupt change boundary. Multiple adjacent annular template regions with a density difference not exceeding the threshold are merged into one feature analysis subdomain to ensure that the density distribution of data points within each subdomain is similar. Ultimately, multiple feature analysis subdomains are generated, and each subdomain corresponds to a continuous density-similar region.

[0063] Steps S2-7 involve smoothing the boundaries of each feature analysis subdomain to generate the final region partitioning result and assigning a unique identifier to each subdomain. Specifically, this includes: smoothing the boundaries of each feature analysis subdomain; selecting multiple consecutive points on the boundary to address potential sharp angles or irregular protrusions; adjusting the coordinates of these points to make the boundary transition smoother; for example, if a segment of the boundary has a distinct polyline, calculating the midpoint between the two endpoints of the polyline and replacing it with a smooth curve from the starting point through the midpoint to the ending point, ensuring that the processed boundary matches the actual distribution trend of the data points and avoiding deviations in subsequent feature analysis due to boundary irregularities; next, verifying the integrity of the smoothed subdomains by checking each... The subdomains are checked to ensure they completely encompass the original similar density data points and that there are no overlaps or gaps between them. If overlaps exist, the boundaries of adjacent subdomains are adjusted; if gaps exist, boundary points are added based on the density characteristics of surrounding subdomains to ensure all data points can be assigned to their corresponding subdomains. Then, a unique identifier is assigned to each feature analysis subdomain. This identifier can be a number or a combination of letters and numbers, assigned sequentially starting from the first subdomain to ensure that each subdomain's identifier is unique and can be associated with the subdomain's location, density characteristics, and other information, facilitating subsequent individual analysis of data points in different subdomains. Finally, the boundary coordinates, number of data points contained within, density range, and unique identifier of each subdomain are recorded to form the final region division result.

[0064] In this embodiment of the invention, the boundary reference points are accurately obtained. Step S2-4 involves emitting probe rays at equal angles along the radial direction from the geometric center point of the convex hull polygon and obtaining the coordinates of the intersection points. This can uniformly cover all directions of the convex hull boundary, ensuring the accuracy of the boundary reference points for the subsequent construction of the annular template. Step S2-5 generates equally spaced segmented points according to a preset ratio and connects them into a closed annular curve to form multiple concentrically distributed reference benchmark annular templates. This allows the templates to cover the internal space of the convex hull layer by layer, avoiding regional overlap or omissions during density statistics. To achieve precise subdomain division based on density, step S2-6 determines abrupt boundary changes by statistically analyzing the density of data points within the annular template and analyzing the density gradient. This divides the region into feature analysis subdomains with similar densities, accurately distinguishing data distribution differences in different icing areas (such as dense and loose icing areas). To improve the practicality and recognizability of subdomains, step S2-7 smooths the boundaries of the feature analysis subdomains to avoid irregular boundaries affecting subsequent feature analysis. A unique identifier is assigned to each subdomain to facilitate individual processing and tracking of data points in different subdomains. This ensures that subsequent error correction, attribute extraction, and other operations can accurately correspond to specific regions, improving the orderliness and accuracy of the overall detection process.

[0065] In a preferred embodiment of the present invention, based on the curvature properties and spatial arrangement of data points within each feature analysis subdomain, a spatial compensation coefficient is calculated using principal component analysis. This spatial compensation coefficient is then used to correct the position parameters of the data points to obtain an optimized processing signal, including:

[0066] Steps S2-8 involve calculating the curvature attribute values ​​of each data point within each feature analysis subdomain, including mean curvature and Gaussian curvature, and analyzing the spatial distribution pattern of the data points. Specifically, this includes: identifying all data points within each feature analysis subdomain; selecting all data points belonging to that subdomain from the set of discrete 3D data points based on the subdomain's boundary coordinates and unique identifier, ensuring no omissions or misselections; then calculating the curvature attribute value of each data point; for a single data point, first selecting a preset number of neighboring data points (e.g., 5-8, adjusted according to the subdomain's data point density); fitting the local surface where the data point is located using the spatial coordinates of these neighboring points; and analyzing the curvature based on the fitted local surface. The curvature of the surface is calculated to obtain the average curvature of the data point, reflecting the overall curvature trend of the surface, and the Gaussian curvature, reflecting the local concavity and convexity characteristics of the surface, ensuring that each data point can obtain two curvature attribute values. Then, the spatial distribution pattern of the data points is analyzed, and the average distance between the neighboring points of all data points in the subdomain is statistically analyzed. If the average distance difference is small and the distribution is uniform, it is determined to be a uniform distribution pattern. If there are some areas where the average distance between neighboring points is significantly smaller, forming dense clusters, it is determined to be an aggregated distribution pattern. If the distance between neighboring points of the data point along a certain direction (such as parallel to the direction of the conductor) is even smaller, showing a strip-shaped distribution, it is determined to be a directional distribution pattern. Finally, the spatial distribution pattern of the data points in each subdomain is determined.

[0067] Steps S2-9 involve calculating the three principal component directions and their corresponding eigenvalues ​​based on the spatial coordinates of data points within each feature analysis subdomain using principal component analysis to determine the main distribution direction of data points within that subdomain. Specifically, this includes: organizing the spatial coordinates of all data points within the target feature analysis subdomain; arranging the X-axis, Y-axis, and Z-axis coordinates of each data point in rows or columns to construct a data point coordinate matrix for that subdomain, ensuring the matrix contains complete coordinate information for all data points within the subdomain; then, performing data centering on the coordinate matrix; calculating the average X-axis coordinate, average Y-axis coordinate, and average Z-axis coordinate of all data points within the subdomain; subtracting the average coordinate from the coordinate value of each data point to obtain the centered coordinate matrix, thus eliminating the influence of the origin position on subsequent analysis. Then, the covariance matrix is ​​calculated based on the centered coordinate matrix. The elements in the covariance matrix reflect the correlation between the coordinates of data points in different coordinate axes, and this matrix reflects the discrete distribution of data points in three-dimensional space. Next, principal component analysis is performed on the covariance matrix to extract three principal component directions, which are the directions corresponding to the three eigenvectors with the largest eigenvalues ​​in the covariance matrix. These three directions represent the three dimensions where the distribution of data points differs most significantly within the subdomain. At the same time, the eigenvalues ​​corresponding to each principal component direction are recorded. The larger the eigenvalue, the higher the degree of dispersion of the data points in that direction. Finally, the three principal component directions are sorted according to the size of their eigenvalues. The principal component direction with the largest eigenvalue is the main distribution direction of the data points in that subdomain, thus determining the core distribution trend of the data points in the subdomain.

[0068] Step S2-10: Based on the main distribution direction of the data points within the subdomain and the curvature attributes of each data point, calculate the spatial compensation coefficient for each data point. Specifically, this includes: determining the weighting coefficient of the main distribution direction; calculating the proportion of the main distribution direction eigenvalue to the sum of the three principal component eigenvalues ​​obtained in step S2-9; this proportion is the weighting coefficient of the main distribution direction. The larger the proportion, the greater the impact of the distribution difference of the data points in that direction on the overall error. Next, determine the curvature influence coefficient based on the curvature attributes of the data points. For data points with a large average curvature or a large absolute value of Gaussian curvature, the local surface curvature is more pronounced, making positional deviations more likely during data acquisition. Therefore, the curvature influence coefficient for these data points is set to a larger value. For data points with smaller curvature... For data points with gently curving surfaces, the curvature influence coefficient is set to a small value to ensure that the influence of curvature on the error is reflected through the coefficient. Then, the deviation of the data points relative to the main distribution direction is calculated. Based on the centered coordinates, the projected distance of each data point in the main distribution direction is calculated. The ratio of the projected distance to the feature value of that direction is the degree of deviation. The greater the degree of deviation, the more obvious the deviation of the data point from the overall trend in the main distribution direction, and the greater the compensation required. Finally, the spatial compensation coefficient of each data point is calculated by weighted summation, combining the weight coefficient of the main distribution direction, the curvature influence coefficient, and the degree of deviation. The greater the weight coefficient, the greater the degree of deviation, and the greater the curvature influence coefficient, the larger the corresponding spatial compensation coefficient value, ensuring that the compensation coefficient can accurately match the error of the data points.

[0069] Step S2-11 involves using the spatial compensation coefficient to perform weighted correction on the spatial coordinates of each data point to generate an optimized processing signal. Specifically, this includes: obtaining the original spatial coordinates and corresponding spatial compensation coefficients of each data point to ensure a one-to-one correspondence without misalignment or matching errors; then, calculating the coordinate correction amount for each data point. Based on the deviation direction of the data point in the main distribution direction (e.g., biased towards the positive or negative direction of the main distribution direction), and combined with the magnitude of the spatial compensation coefficient, the direction and magnitude of the coordinate correction are determined. If the data point is biased towards the positive direction of the main distribution direction and the deviation is large, and the compensation coefficient is large, a larger correction amount is applied in the negative direction; if it is biased towards the negative direction and the deviation is small, and the compensation coefficient is small, a smaller correction amount is applied in the positive direction. Simultaneously, the curvature attributes of the data points are referenced, and the coordinates of data points in areas with large curvature are further fine-tuned to ensure that the corrected coordinates better match the actual icy surface morphology. Then, the original spatial coordinates of the data points are weighted and corrected. The calculated coordinate correction is superimposed on the corresponding original coordinate values ​​to obtain the corrected X-axis, Y-axis, and Z-axis coordinates of each data point, thus completing the position parameter error correction for a single data point. Finally, all corrected data points within the feature analysis subdomains are integrated and classified according to the unique identifier of the subdomain to form a set containing the spatial coordinates of all corrected data points. This set is the optimized processing signal, which can effectively reduce the errors caused by the discreteness, curvature deviation, and uneven distribution of the original data.

[0070] In this embodiment of the invention, steps S2-8 calculate the average curvature and Gaussian curvature of data points in each subdomain and analyze the spatial distribution pattern. This clearly identifies the bending characteristics of the icing surface and the dense / sparse areas of data, providing a targeted basis for subsequent error correction and avoiding correction deviations caused by ignoring local morphological differences. Step S2-9 uses principal component analysis to determine the main distribution direction of data points within the subdomain, accurately locating the dimension with the most significant data dispersion. This allows the compensation coefficient calculation to focus on the direction with the greatest impact on the error, improving the targeting and efficiency of subsequent correction. Step S2-10 combines the main distribution direction and curvature attributes to calculate the compensation coefficient, considering both the overall data distribution trend and the errors caused by local surface curvature, ensuring that the compensation coefficient for each data point accurately matches its error level. Step S2-11 uses the compensation coefficient to weight and correct the coordinates of the data points, effectively eliminating the positional errors caused by the dispersion, spatial arrangement deviation, and surface curvature of the original data. The generated optimized signal can more realistically reflect the actual spatial morphology of icing on transmission lines.

[0071] In a preferred embodiment of the present invention, step S3, which involves obtaining a clean transmission signal by eliminating transmission interference during channel switching using the Manchester coding protocol and a local buffer verification algorithm, includes:

[0072] The optimized signal is preprocessed using Manchester encoding. First, the structure of the optimized signal is analyzed. This signal contains a set of coordinates of three-dimensional discrete data points corrected for each feature analysis subdomain, arranged in order of subdomain unique identifiers and data point timestamps. According to the Manchester encoding protocol, each data bit in the signal is converted into an encoded signal with clock synchronization information. Specifically, within the transmission period of each data bit, a level transition is set at the middle of the period to distinguish the data value by the direction of the transition: a transition from low to high level represents "0", and a transition from high to low level represents "1". This encoding method enables the signal to have built-in clock synchronization capability, avoiding data decoding errors caused by clock offsets during subsequent transmission. It also enhances the signal's resistance to low-frequency interference in the channel, ensuring that the encoded signal maintains transmission stability in the complex electromagnetic environment of power transmission lines.

[0073] Next, signal transmission control is implemented during the channel switching process. First, the quality of the current transmission channel is monitored in real time. By detecting three key indicators—signal strength, bit error rate, and transmission delay—the stability of the channel is determined. If the signal strength is below a preset threshold (adapting to signal attenuation characteristics in high-altitude areas), the bit error rate exceeds 0.1%, or the transmission delay is greater than 500 milliseconds, the current channel is deemed to have interference risk, triggering the channel switching mechanism. Before switching, a backup channel connection is established in advance. Based on the multi-channel list preset by the transmission line monitoring system (including wired and wireless channels, such as 4G / 5G or dedicated power communication channels), the backup channel with the best current quality is selected. A handshake protocol is used to synchronize the backup channel between the sending and receiving ends, ensuring uninterrupted data transmission during switching. During switching, a seamless switching strategy of transmitting first and then disconnecting is adopted. While the backup channel begins transmitting coded signals, the current channel transmission is maintained until it is confirmed that the backup channel is receiving data normally before disconnecting the current channel, avoiding data loss or duplicate transmissions during switching.

[0074] Then, a local cache verification algorithm is activated to verify data integrity. Local cache modules are set up at both the sending and receiving ends. The storage capacity of the cache modules is set according to the transmission rate of the optimized signal (matching a data acquisition frequency of 3-5 times per second to ensure at least 10 seconds of encoded signal data can be temporarily stored). While transmitting the encoded signal, the sending end stores the signal data in timestamp segments in the local cache and generates a unique CRC checksum for each segment (containing data length, timestamp, and data characteristic value), which is transmitted to the receiving end along with the data. Upon receiving the data, the receiving end first temporarily stores the data in the local cache, then extracts the attached CRC checksum and performs a verification calculation on the received data. If the calculation result matches the received checksum, the data is considered complete, marked as valid, and used for subsequent processing. If the verification result does not match, the data is considered to have an error due to interference, and a retransmission request is sent to the sending end. The request includes the timestamp and segment identifier of the erroneous data. The sending end retrieves the original encoded data of the corresponding segment from the local cache based on the identifier and retransmits it until the receiving end's verification passes.

[0075] Finally, secondary filtering of interference signals and generation of clean signals are performed. The receiving end performs Manchester decoding on the verified encoded signal, and restores the original optimized signal data structure by identifying the level transitions in the middle of each data bit. After decoding, interference filtering is performed on the signal: a sliding window filtering method is used, with 10 data points as a window, and the average coordinates of the data points in the window are calculated. Abnormal data points that deviate from the average value by more than a preset range (set according to the normal dispersion of icing data points) are removed. These abnormal points are usually caused by sudden electromagnetic interference. After filtering, the remaining valid data points are reordered according to the unique identifier of the subdomain and the timestamp, and integrated into a continuous and error-free signal set, which is the clean transmission signal. After generation, a final integrity check is performed on the clean transmission signal to confirm that the number of data points is consistent with the number of the original optimized signal at the transmitting end, and that the spatial coordinate range of each subdomain data conforms to the actual monitoring area of ​​the transmission line, ensuring that the clean transmission signal can be directly used for subsequent icing analysis by deep convolutional neural networks.

[0076] In a preferred embodiment of the present invention, step S4 involves inputting the clean transmission signal into a pre-trained deep convolutional neural network, extracting the spatial attributes of the ice layer through multiple convolutional and pooling layers, and then performing feature fusion through a fully connected layer to calculate the ice layer thickness distribution, including:

[0077] Step S4-1 converts the clean transmission signal into a three-dimensional feature tensor. Specifically, this includes: parsing the structure of the clean transmission signal, which contains the three-dimensional spatial coordinates of data points after error correction and corresponding feature information (such as subdomain identifiers, density features, etc.), which needs to be organized into an ordered set according to the spatial positional relationship of the data points; next, determining the dimensional framework of the three-dimensional feature tensor, using the spatial range of the transmission line icing monitoring area as a benchmark, setting the X and Y axes as the planar dimension, and the Z axis as the height dimension, while using the curvature attributes and density features of the data points as channel dimensions, forming a three-dimensional structure of planar size × height size × number of channels; then, filling the tensor data, mapping the feature information of each data point to the corresponding position in the three-dimensional tensor according to its spatial coordinates; for positions without data points within the coordinate range, using the feature mean of neighboring valid data points to fill, ensuring the tensor has no empty values; finally, normalizing the tensor, adjusting the feature values ​​of each channel to the 0-1 range, eliminating the influence of differences in feature magnitudes on the neural network, and ultimately forming a three-dimensional feature tensor that meets the input requirements of a deep convolutional neural network.

[0078] Step S4-2: Based on the three-dimensional feature tensor, primary spatial features are extracted through the first convolutional layer, and the extracted primary spatial features are dimensionality-reduced using the first pooling layer to obtain a primary feature map containing key feature information. Specifically, this includes: configuring the parameters of the first convolutional layer, selecting a three-dimensional convolutional kernel of a preset size (e.g., 3×3×3), setting the number of convolutional kernels according to the primary feature extraction requirements (usually 32-64), and setting the stride to 1 to ensure that the convolution operation can cover each local region of the three-dimensional feature tensor. Next, feature extraction is performed through the first convolutional layer. The convolutional kernel slides on the three-dimensional feature tensor, performing weighted summation and nonlinear activation (e.g., ReLU activation function) on the features of each local region to capture primary spatial features such as edges, inflection points, and basic contours of the ice surface, generating an initial feature map containing multiple channels. Then, the first pooling layer is set to max pooling, with a pooling window size of 2×2×2 and a stride of 2, to perform dimensionality reduction on the initial feature map, retaining the maximum feature value within each window, removing redundant information, and enhancing the robustness of the features to slight positional shifts. Finally, the output is a dimensionality-reduced primary feature map, which retains the key primary features while reducing its size to half that of the input.

[0079] Step S4-3: Based on the primary feature map, intermediate spatial features are extracted through a second convolutional layer, and the intermediate spatial features are compressed using a second pooling layer to generate an intermediate feature map with enhanced expressive power. Specifically, this includes configuring the parameters of the second convolutional layer, selecting more three-dimensional convolutional kernels (e.g., 64-128) than the first convolutional layer, while maintaining the kernel size at 3×3×3 and a stride of 1, to extract more complex intermediate spatial features. Next, the primary feature map is processed through depth using the second convolutional layer. The convolutional kernels combine edge and contour information from the primary features to further extract intermediate features such as regional textures of the ice layer, local convex / concave morphology, and boundary transitions of different icing areas, generating an initial intermediate feature map with more channels. Then, the second pooling layer is set to average pooling, with a pooling window size of 2×2×2 and a stride of 2, to compress the initial intermediate feature map. By calculating the average value of the features within the window, the feature dimension is reduced, while smoothing feature fluctuations and enhancing the ability to suppress local noise. Finally, an intermediate feature map is generated, which is further reduced to half the size of the primary feature map. It has a stronger feature representation ability and can more clearly reflect the medium-scale spatial attributes of the ice layer.

[0080] Step S4-4: Based on the intermediate feature map, high-level spatial features are extracted through a third convolutional layer to capture detailed information about the ice layer distribution and generate a high-level feature map. Specifically, this includes configuring the parameters of the third convolutional layer, selecting more 3D convolutional kernels (e.g., 128-256), adjusting the kernel size to 3×3×3 or 5×5×5, and setting a stride of 1 to capture finer spatial details. Next, the intermediate feature map is deeply mined through the third convolutional layer. The convolutional kernels combine regional texture and morphological information from the intermediate features to extract key high-level spatial features such as abrupt changes in ice thickness, the boundary between ice cover and guide lines, and the distribution details of locally dense ice cover. These features are directly related to the spatial variation patterns of ice thickness. During the extraction process, no pooling layers are used to preserve the dimensional details of the feature map and ensure that the fine spatial information required for thickness calculation is not lost. Finally, a high-level feature map is generated, with the same size as the intermediate feature map, a further increased number of channels, and richer feature information, accurately reflecting the microscopic differences in the ice layer distribution.

[0081] Steps S4-5 involve flattening the high-level feature map into a one-dimensional feature vector and inputting it into a fully connected layer for feature fusion and transformation to obtain the fused features. Specifically, this includes: flattening the high-level feature map into a one-dimensional feature vector; sequentially reading the feature values ​​of each spatial location in the high-level feature map according to channel order; converting the three-dimensional structure (length × width × height × number of channels) into a one-dimensional array, with the array length being the product of the dimensions, ensuring that all high-level feature information is completely preserved. Next, a fully connected layer structure is configured, with 2-3 fully connected layers. The number of neurons in the first layer is 1 / 4 to 1 / 2 of the length of the flattened feature vector, and the number of neurons in subsequent layers gradually decreases. The number of neurons in the last layer is set according to the fusion requirements (e.g., 512 or 1024). Then, the one-dimensional feature vector is input into the fully connected layer, and cross-channel and cross-space feature fusion is achieved through the weight connection between neurons. Key information (such as contour, texture and details) in primary, intermediate and high-level features is combined, and feature transformation is performed through non-linear activation function (such as ReLU) to enhance the non-linear expressive ability of the features. Finally, the fused feature vector is output, which integrates all the key spatial attributes of the ice layer from macro to micro.

[0082] Steps S4-6 involve calculating the ice thickness value corresponding to each spatial location based on the fused features using a regression output layer to generate an ice thickness distribution map. Specifically, this includes configuring the regression output layer, setting the number of neurons corresponding to the spatial size of the monitoring area, with each neuron outputting the ice thickness at a specific spatial location, and using a linear activation function (such as Identity) to ensure the output value covers the actual ice thickness range (e.g., 0-100mm). Next, the fused feature vector is input into the regression output layer, and the predicted ice thickness value for each spatial location is calculated using neuron weights. The system optimizes the ice thickness data from the pre-training process to ensure minimal deviation between the predicted and actual thicknesses. Then, it maps the output thickness predictions to corresponding spatial coordinates, assigning each prediction to its corresponding X, Y, and Z coordinates according to the original spatial dimensions of the 3D feature tensor, thus forming preliminary ice thickness distribution data. Finally, it smooths the preliminary distribution data by using Gaussian filtering to eliminate local prediction noise, ensuring a natural transition between thickness values ​​at adjacent spatial locations. This results in an ice thickness distribution map that clearly reflects the icing thickness in different areas of the transmission line, demonstrating the spatial differences in ice thickness.

[0083] In this embodiment of the invention, the clean transmission signal is converted into a three-dimensional feature tensor. Through data structure standardization, gap filling, and normalization, the data conforms to the input format of a deep convolutional neural network, while eliminating interference from differences in feature magnitude. Step S4-2 captures primary spatial features such as ice layer edges and basic contours through a first convolutional layer, followed by dimensionality reduction through a first pooling layer. This retains key information while eliminating redundant data, reducing subsequent computational load, improving network processing efficiency, and enhancing the robustness of features to slight positional shifts. Step S4-3 uses a second convolutional layer to extract intermediate features such as ice layer texture and local morphology, combined with average pooling to compress the data. This smooths feature fluctuations, suppresses noise, and further enhances the features of medium-scale features within the ice layer. The ability to express spatial attributes; Step S4-4 uses the third convolutional layer to mine high-level detailed features such as ice thickness abrupt change points and the boundary between icing and conductors, without pooling to preserve spatial size details and ensure that key microscopic information is not lost; Step S4-5 flattens the high-level feature map and inputs it into the fully connected layer to achieve cross-channel and cross-space fusion from primary to high-level features, integrating the macroscopic outline and microscopic details of the ice layer to form a more comprehensive feature vector; Step S4-6 calculates the ice thickness at each spatial location through the regression output layer, and generates a thickness distribution map after smoothing, which can intuitively and quantitatively display the differences in ice thickness in different areas of the transmission line, providing accurate quantitative basis for operation and maintenance personnel to judge icing risks and formulate de-icing plans.

[0084] In a preferred embodiment of the present invention, an ice layer category is identified using a softmax classifier. Based on the thickness distribution map and the ice layer category identification result, a detection result containing ice layer thickness values ​​and spatial distribution attributes is generated, including:

[0085] Steps S4-7 involve extracting a feature vector for classification based on the fused features output by the fully connected layer, and inputting this feature vector into a softmax classifier. Specifically, this includes: analyzing the composition of the fused features output by the fully connected layer. These fused features contain multiple feature components, which comprehensively reflect information such as the macroscopic outline of the ice layer (e.g., overall shape boundary), intermediate texture (e.g., surface roughness), microscopic details (e.g., local protrusions and depressions), and spatial variation patterns related to thickness. Dimensions directly related to ice layer category identification are selected from these feature components. Specifically, this involves analyzing the correlation between each feature component and known ice layer categories, retaining features that show significant differences between different ice layer categories. For example, internal structure feature components reflecting ice layer density (different ice layer categories have significantly different densities; rime ice has a lower density, while glaze ice has a higher density); feature components reflecting surface texture (rime ice surfaces are often velvety, while glaze ice surfaces are relatively smooth); feature components reflecting local curvature (different ice layer categories have different degrees of surface curvature due to different formation processes); and feature components related to reflectivity (different ice layer categories have different reflectivity to radar or ultrasound). At the same time, feature components that are only related to the calculation of ice thickness are removed, such as features that simply reflect changes in thickness. These features do not help to distinguish ice types. This screening ensures that the remaining features can effectively distinguish different ice types such as rime ice, rain ice, and mixed ice.

[0086] Next, adjust the dimension of the filtered feature vector according to the preset number of ice categories. The preset number of categories is usually determined based on the common icing types of transmission lines, generally 3 to 5 categories. If the preset number is 3, namely hoarfrost ice, rain ice, and mixed ice, then the dimension of the filtered feature vector needs to be adjusted to a length suitable for a 3-class classification task. If the preset number is 5, adding categories such as wet snow ice and frost ice, then adjust it to a length suitable for a 5-class classification task. During the adjustment process, the integrity of the feature vector must be ensured, without losing the key feature information selected. It is only through feature recombination or dimension transformation that it is made to match the input requirements of the softmax classifier, ensuring that the classifier can successfully receive and process the feature vector.

[0087] Then, the adjusted feature vectors are normalized. Since the numerical ranges of different feature components may vary greatly, some feature components may have large values ​​and some may have small values. This difference in magnitude will affect the softmax classifier's weight allocation of features, causing feature components with large values ​​to have too much influence on the classification result, while the role of feature components with small values ​​will be masked. The normalization process compresses the values ​​of all feature components into the range of 0 to 1. Specifically, the maximum and minimum values ​​of each feature component are calculated, and then each feature value is subtracted from the minimum value and divided by the difference between the maximum and minimum values. This process ensures that the values ​​of each feature component are of the same magnitude, eliminating the interference of magnitude differences on the classification result, and allowing each feature component to make a balanced contribution in the classification process.

[0088] Finally, the normalized classification-specific feature vectors are input into a pre-trained softmax classifier. This classifier has been trained on a large amount of labeled ice layer category sample data and is capable of recognizing the feature patterns of different ice layer categories. After the feature vectors are input, the classifier will perform calculations based on these features, laying the foundation for subsequent calculations of the probability distribution of ice layer categories. During the input process, it is crucial to ensure the complete transmission of the feature vectors, without data loss or misalignment, to guarantee that the classifier can accurately receive and use this feature information.

[0089] The specific implementation process of the pre-trained softmax classifier is as follows:

[0090] The specific implementation process of the pre-trained softmax classifier is as follows:

[0091] First, a training dataset for ice categories is constructed by collecting a large number of sample data containing different ice categories. These samples need to cover common icing types of transmission lines, such as rime ice, rain ice, and mixed ice. Each sample should contain a corresponding classification-specific feature vector and a clear ice category label. The feature vectors need to be obtained through the same extraction process as the actual detection to ensure that they are consistent with the feature format input to the classifier. The collected sample data is cleaned to remove invalid samples with missing features or incorrect labels, and data augmentation techniques are used, such as making small perturbations to the feature vectors and generating new samples, to expand the dataset size and reduce the risk of overfitting. Finally, the processed dataset is divided into training and validation sets according to a preset ratio, usually with the training set accounting for 70% to 80% and the validation set accounting for 20% to 30%.

[0092] Next, the network structure of the softmax classifier is designed. The classifier adopts a fully connected network form, and the number of neurons in the input layer is consistent with the dimension of the classification-specific feature vector to ensure that the feature vector can be completely input. One to two hidden layers are set in the middle, and the number of neurons in the hidden layer is set according to the feature complexity, usually 1 / 2 to 2 / 3 of the number of neurons in the input layer. A ReLU activation function is added after the hidden layer to enhance the non-linear expressive ability of the network. The number of neurons in the output layer is the same as the preset number of ice layer categories, and each neuron corresponds to one ice layer category. The softmax function is connected after the output layer to convert the neuron output into the probability distribution of each ice layer category.

[0093] Next, configure the training parameters and loss function; select the Adam optimizer, set the initial learning rate to 0.001, and dynamically adjust it according to the loss changes during training. The learning rate adjustment strategy adopts a step decay, and the learning rate is halved when the validation set loss does not decrease for several consecutive rounds; the cross-entropy loss function is used to measure the difference between the predicted class probability distribution and the true label; the training batch size is set to 32 or 64, depending on the hardware computing power; the total number of training rounds is set to 100 to 200 rounds, and an early stopping mechanism is set. If the validation set loss does not decrease significantly within 30 rounds, training is terminated early to avoid overtraining.

[0094] Next, the model pre-training process is performed. First, the training set data is trained iteratively. In each iteration, a batch of samples is input into the classifier. The class probability distribution of each sample is calculated through forward propagation. The prediction loss is calculated using the loss function. Then, the weight parameters of each layer of the network are updated through the backpropagation algorithm to continuously reduce the loss value. After each training round, the model performance is evaluated using the validation set, the classification accuracy is calculated, and the loss value and accuracy of the current round are recorded. During the training process, the model parameters with the highest accuracy on the validation set are saved as the candidate optimal model.

[0095] After training, the pre-trained model is evaluated and optimized. The performance of the saved optimal model is tested using an independent test set, and the classification accuracy, precision, recall and other metrics on the test set are calculated. If the performance on the test set does not meet expectations, the reasons are analyzed and optimizations are made, such as adjusting the number of hidden layers and neurons in the network, changing the optimizer or adjusting the learning rate, increasing the number of training samples, etc. The training and evaluation process is repeated until the classification accuracy of the model on the test set reaches the preset threshold, which is usually not less than 90%.

[0096] Finally, the pre-trained model parameters are saved, and the optimized model weights, biases, and other parameters are stored in file form to ensure the integrity and readability of the parameters. In practical applications, these pre-trained parameters are directly loaded to initialize the softmax classifier, which can then receive the input classification-specific feature vectors and quickly output the probability distribution of ice layer categories, thereby achieving accurate identification of ice layer categories.

[0097] Steps S4-8 involve processing the input feature vector using the softmax classifier to calculate the probability distribution of each ice layer category, and selecting the category with the highest probability as the final identification result for the ice layer category. Specifically, this includes:

[0098] The first stage involves determining the pre-defined ice layer category system for the softmax classifier:

[0099] This study analyzes the differences between the actual physical characteristics of icing on transmission lines and the severity of hazards in operation and maintenance scenarios, and classifies ice layers accordingly. Furthermore, it clarifies the classification criteria by considering the icing formation mechanisms in typical transmission line environments such as high altitude and low temperature: ice layers with a density less than 0.3 g / cm³ are classified. 3 Ice with a loose structure that easily detaches under wind force is defined as rime ice; ice with a density greater than 0.8 g / cm³ is considered rime ice. 3 Ice that has a smooth and hard surface, strong adhesion to conductors, and is prone to causing line overload or breakage is defined as rime ice; ice that is formed by the alternating freezing of frost ice and rime ice under fluctuating temperature conditions and has an uneven structure is defined as mixed ice; ice with a moisture content of more than 30% that is easily frozen into blocks due to low temperature is defined as wet snow ice.

[0100] Next, a unique label is assigned to each category. The label is in the form of a numerical code (e.g., 0 for rime ice, 1 for rain ice, 2 for mixed ice, and 3 for wet snow ice) to ensure that the label corresponds one-to-one with the category and remains consistent in subsequent probability calculations and result output. At the same time, the core features of each category (e.g., "low density, easy to fall off" for rime ice and "high density, high adhesion" for rain ice) are associated with the label and stored as the learning basis for the classifier pre-training stage, and also to provide a reference for maintenance personnel to interpret the recognition results.

[0101] In the second stage, the probability distribution of ice layer categories is calculated using a softmax classifier:

[0102] First, the normalized classification-specific feature vectors are input into the input layer of the softmax classifier. Each dimension in the feature vector (such as features reflecting density and features reflecting surface texture) is matched with the weight parameters learned during the pre-training stage of the classifier.

[0103] Then, the classifier calls upon the feature patterns of different ice layer categories stored during pre-training (such as the combination of high reflectivity and low curvature fluctuation features corresponding to rime ice) to perform a preliminary evaluation of the matching degree between the current input feature vector and each category. During the evaluation process, the matching degree is enhanced by an exponential function: if the matching degree between a certain category and the feature vector is higher than that of other categories, the exponential function will amplify the matching value corresponding to that category, highlighting the contribution of highly correlated features; if the matching degree is low, the corresponding value will be reduced to decrease the interference of low-correlated features.

[0104] Finally, the values ​​of all categories after being processed by the exponential function are normalized: the value corresponding to each category is divided by the sum of the values ​​of all categories, so that the final value of each category (i.e., the probability value) is in the range of 0-1, and the sum of the probability values ​​of all categories is equal to 1, forming a complete probability distribution of ice layer categories (e.g., rime ice probability 0.1, rain ice probability 0.8, mixed ice probability 0.08, wet snow ice probability 0.02).

[0105] The specific calculation formula for the exponential function is as follows:

[0106] Suppose the output layer of the softmax classifier has Number of neurons (corresponding to) (Ice layer category), the original output value obtained after processing the feature vector of the input classifier through a fully connected layer is (in, (where represents the original matching degree corresponding to the i-th category), then the probability value of the i-th ice layer category. The calculation formula is:

[0107] ;

[0108] in, K is a natural constant (approximately equal to 2.71828), which is the base of the exponential function; K represents the total number of preset ice layer categories, which is determined by the actual scenario of power transmission line icing. For example, the "rime ice, rain ice, mixed ice, wet snow ice" mentioned above correspond to K=4. If frost ice categories are added later, then K=5. Its value is completely consistent with the number of neurons in the output layer of the classifier. This represents the initial matching degree of the i-th ice layer category, which is the preliminary processing result of the fully connected layer of the classifier on the input feature vector:

[0109] The fully connected layer calls upon the feature patterns of the i-th type of ice layer stored during the pre-training phase (such as the high reflectivity and low curvature fluctuation feature weights of rime ice), multiplies the values ​​of each dimension of the input feature vector with their corresponding weights, sums them, and then adds the bias term to obtain the final result. ; The larger the value, the higher the matching degree between the current input feature vector and the feature pattern of the i-th type of ice layer; the smaller the value, the lower the matching degree. For example, if the input feature vector conforms to the high reflectivity and low curvature characteristics of rime ice, the rime ice corresponding to... It will be significantly larger than other categories; Indicates the original matching degree The strengthening factor; It is an index variable whose function is to iterate through each of the K types of ice layers.

[0110] The third stage involves confidence verification and category label extraction of the probability distribution results:

[0111] First, extract the maximum probability value from the generated category probability distribution, and record the ice layer category label corresponding to the maximum probability value. For example, if the probability value of 0.8 for rime ice in the probability distribution is the maximum value and the corresponding label is 1, then it is temporarily set as the current category result to be confirmed.

[0112] Next, a preset confidence threshold is invoked (the threshold is set to 0.7 based on historical detection data and operational experience; this threshold can be fine-tuned in different environments), and the extracted maximum probability value is compared with the threshold. If the maximum probability value is higher than 0.7 (such as 0.8 for rime ice mentioned above), it is determined that the current feature vector matches the category sufficiently, and the corresponding category label (such as label 1) is directly extracted as a candidate result; if the maximum probability value is lower than 0.7 (such as 0.65 in a certain detection scenario), the result is marked as a category pending confirmation, and a subsequent review prompt mechanism is triggered (such as recording the time, location, and feature vector of the detection sample in the system background, prompting operations personnel to further verify through manual inspection or supplementary data collection).

[0113] The fourth stage involves determining the final identification result and recording the probability value:

[0114] First, the candidate results that have passed the confidence test are finally confirmed: if the candidate result is not marked as a category to be confirmed, the corresponding category label (such as the rime ice corresponding to label 1) is used as the final identification result of the ice layer category; if it is marked as a category to be confirmed, the final result is not output for the time being, and the result is updated based on the verification data after the operation and maintenance personnel review it.

[0115] Then, record the probability value corresponding to the final identification result (e.g., 0.8 for rime ice), and store the probability value in association with the identification result, detection time, and coordinates of the monitoring section. This probability value will serve as the basis for the credibility of subsequent result analysis: if the probability value is close to 1 (e.g., 0.95), it indicates that the classifier has extremely high confidence in the judgment of the result and can be directly used for operation and maintenance decisions; if the probability value is slightly higher than the threshold (e.g., 0.72), it prompts operation and maintenance personnel to pay attention to the subsequent changes in icing in the area to avoid misjudgment due to feature similarity.

[0116] Finally, the final recognition results and probability values ​​are organized according to a preset format.

[0117] Steps S4-9 involve performing spatial alignment and feature fusion processing based on the ice thickness distribution map generated by the regression output layer and the ice category identification results obtained by the softmax classifier to obtain fused data containing thickness and category information. Specifically, this includes: determining the benchmark for spatial alignment using the spatial coordinate system of the ice thickness distribution map (X-axis parallel to the conductor direction, Y-axis perpendicular to the horizontal direction of the conductor, Z-axis vertical direction) as a unified benchmark; extracting the coordinate information (X, Y, Z) and corresponding ice thickness values ​​for each spatial location in the thickness distribution map to form a coordinate and thickness association dataset; then, associating the ice category identification results with spatial locations. If the category identification result is a single category for the entire area (e.g., the entire monitoring section is rime ice), the category label is mapped to all spatial locations in the thickness distribution map; if the category... If the identification result is a partition category (e.g., the upper part of the monitoring section is rime ice, and the lower part is mixed ice), then based on the spatial range of the sub-domains analyzed in the previous steps, the category label of each sub-domain is mapped to all spatial locations within the sub-domain, ensuring that each coordinate and thickness data matches the corresponding category label. Then, feature fusion processing is performed, integrating the coordinates, thickness values, and category labels of each spatial location into a fused data record of coordinates, thickness, and category. At the same time, the data alignment accuracy is checked by randomly sampling some spatial locations to verify whether their thickness values ​​and category labels match the actual icing characteristics (e.g., the thickness of the area corresponding to rime ice is usually greater and the density is higher). If there is a misalignment, the association between the category and the spatial location is readjusted. Finally, a complete fused dataset is generated, which contains the thickness and category information of all spatial locations in the monitoring area.

[0118] Step S4-10: Based on the fused data, generate detection results containing ice thickness values, spatial distribution attributes, and category information. Specifically, this includes: organizing the structure of the fused dataset; grouping the fused data according to the actual monitoring sections of the transmission line (e.g., the conductor segment between tower A and tower B, the section with the maximum conductor sag, and the hardware section connecting the tower and the conductor); arranging data records within each segment sequentially along the Z-axis (vertical direction) or Y-axis (horizontal direction) of spatial coordinates to clearly present the icing situation in different segments; then, extracting the ice thickness values ​​for each segment, calculating the maximum, minimum, and average thickness for each segment, and marking the spatial locations of thickness abrupt changes (e.g., the thickness at X=100m, Y=5m in a certain segment suddenly increases from 10mm to 25mm), quantifying the severity of icing with specific values; and then supplementing the spatial distribution... The system describes the spatial distribution characteristics of ice in each section, combining thickness data and coordinate information. For example, the ice thickness in the upper part (Z>15m) of the conductor in section AB of the tower is generally greater than that in the lower part (Z<15m), with the maximum thickness concentrated near the lowest point of the sag. It also marks the distribution areas of different ice types, such as rime ice mainly distributed in the lower half of section AB and hoarfrost ice in the upper half. Finally, all information is integrated to generate the detection results, which include the detection time, monitoring section number, thickness statistics for each section (maximum / minimum / average thickness), spatial distribution description (thickness differences and location correlation), ice type and distribution range, and a simplified text description of the thickness distribution map, such as a section where the thickness is thicker in the middle and thinner at both ends. This ensures the detection results are intuitive and easy to understand, facilitating maintenance personnel to quickly locate high-risk icing sections.

[0119] In this embodiment of the invention, step S4-7 extracts feature vectors strongly correlated with ice layer categories from the fusion features of the fully connected layer, removes redundant information, and adapts to the input requirements of the softmax classifier to ensure that the features input to the classifier can effectively distinguish different ice layer types (such as rime ice and rain ice); step S4-8 calculates the probability distribution of each category through the softmax classifier, selects the category with the highest probability as the recognition result, and ensures the reliability of the result by relying on confidence verification to avoid fuzzy classification; a clear ice layer category (such as high-density rain ice) can provide a key basis for assessing the degree of icing hazard (such as rain ice with strong adhesion and greater hazard), making up for the limitations of judging solely by thickness; step S4-9 The thickness distribution map and category identification results are spatially aligned to ensure that the thickness value at each spatial location matches the corresponding ice layer category, avoiding information misalignment. Complete data of coordinates, thickness, and category is formed through feature fusion to ensure that no ice-related information is omitted. Steps S4-10 integrate and fuse the data to output detection results that include thickness values ​​(maximum / minimum / average thickness), spatial distribution attributes (such as the maximum thickness at sag), and category information. This not only quantifies the ice thickness but also clearly presents the spatial distribution pattern of thickness and category, making it easier for maintenance personnel to quickly locate high-risk sections (such as tower connection sections covered by rime ice with a thickness exceeding the threshold) and develop targeted de-icing and protection plans to improve maintenance efficiency and accuracy.

[0120] like Figure 2 As shown, an embodiment of the present invention provides a transmission line icing detection system, comprising:

[0121] The acquisition module is used to acquire raw echo data from the surface of the transmission line and to preprocess the raw echo data to obtain a preprocessed signal.

[0122] The correction module is used to convert the preprocessed signal into a set of discrete 3D data points, extract the convex hull boundary of the discrete 3D data point set based on the Graham scan algorithm, and determine the outer morphological features of the data points; establish multiple reference ring templates within the outer morphological features using a midpoint closed curve generation algorithm, divide the region according to the density distribution characteristics of the data points, and generate multiple feature analysis subdomains; calculate the spatial compensation coefficients using principal component analysis based on the curvature attributes and spatial arrangement of the data points in each feature analysis subdomain, and use the spatial compensation coefficients to correct the position parameters of the data points to obtain an optimized processed signal;

[0123] The elimination module is used to eliminate transmission interference and obtain a clean transmission signal during channel switching by using the Manchester coding protocol and local buffer verification algorithm.

[0124] The prediction module is used to input the clean transmission signal into a pre-trained deep convolutional neural network, extract the spatial attributes of the ice layer through multiple convolutional and pooling layers, perform feature fusion through a fully connected layer, calculate the ice layer thickness distribution, identify the ice layer category through a softmax classifier, and generate a detection result containing the ice layer thickness value and spatial distribution attributes based on the thickness distribution map and the ice layer category identification result.

[0125] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for detecting icing on power transmission lines, characterized in that, The method includes: Raw echo data from the surface of the transmission line is collected, and the raw echo data is preprocessed to obtain a preprocessed signal. The preprocessed signal is converted into a set of discrete 3D data points. The Graham scan algorithm is used to extract the convex hull boundary of the set of discrete 3D data points to determine the peripheral morphological features of the data points. A midpoint closed curve generation algorithm is used to establish multiple reference ring templates within the peripheral morphological features. Based on the density distribution characteristics of the data points, the regions are divided to generate multiple feature analysis subdomains. Based on the curvature attributes and spatial arrangement of the data points in each feature analysis subdomain, principal component analysis is used to calculate spatial compensation coefficients. These spatial compensation coefficients are then used to correct the positional parameters of the data points to obtain an optimized processed signal. Specifically, a midpoint closed curve generation algorithm is used to establish multiple reference ring templates within the outer morphological features. Based on the data point density distribution characteristics, the region is divided to generate multiple feature analysis subdomains, including: Based on the geometric center point of the convex hull polygon structure, probe rays are emitted radially at equal angles to the boundary of the convex hull, and the coordinates of the intersection points of each ray and the boundary are obtained. Using the geometric center point as a reference, multiple equally spaced segment points are generated on each ray according to a preset ratio. The segment points of the same ratio on adjacent rays are connected to form a closed loop curve, and multiple concentrically distributed reference loop templates are constructed. The density distribution of data points inside each annular template is statistically analyzed, density gradient analysis is used to determine the density abrupt boundary, and the entire region is divided into multiple feature analysis subdomains with similar density characteristics based on the density distribution characteristics. Boundary smoothing is performed on each feature analysis subdomain to generate the final region partitioning result, and a unique identifier is assigned to each subdomain; Specifically, based on the curvature attributes and spatial arrangement of data points within each feature analysis subdomain, spatial compensation coefficients are calculated using principal component analysis. These spatial compensation coefficients are then used to correct errors in the position parameters of the data points, resulting in an optimized processing signal. For each feature analysis subdomain, calculate the curvature attribute values ​​of each data point within the domain, including average curvature and Gaussian curvature, and analyze the spatial distribution pattern of the data points; Based on the spatial coordinates of data points within each feature analysis subdomain, the principal component analysis method is used to calculate the directions of the three principal components and their corresponding eigenvalues, thereby determining the main distribution direction of the data points within that subdomain. Calculate the spatial compensation coefficient for each data point based on the main distribution direction of the data points in the subdomain and the curvature properties of each data point; The spatial coordinates of each data point are weighted and corrected using the aforementioned spatial compensation coefficients to generate an optimized processing signal; The optimized signal is processed by using the Manchester coding protocol and local buffer verification algorithm during channel switching to eliminate transmission interference and obtain a clean transmission signal. The clean transmission signal is input into a pre-trained deep convolutional neural network. The spatial attributes of the ice layer are extracted through multiple convolutional and pooling layers, and then the features are fused through a fully connected layer to calculate the ice layer thickness distribution. The ice layer category is identified by a softmax classifier. Based on the thickness distribution map and the ice layer category identification results, a detection result containing the ice layer thickness value and spatial distribution attributes is generated.

2. The method for detecting icing on transmission lines according to claim 1, characterized in that, The preprocessed signal is converted into a set of discrete 3D data points. The Graham scan algorithm is used to extract the convex hull boundary of the discrete 3D data points to determine the peripheral morphological features of the data points, including: The preprocessed signal is converted into a set of discrete data points in three dimensions using a three-dimensional coordinate mapping function, where each data point contains spatial coordinate information; The discrete data point set of the three-dimensional array is sorted by polar angle, with the lowest point as the starting point. All data points are traversed based on the Graham scan algorithm. The convexity of the point set is determined by the cross product of vectors, and the convex hull boundary is gradually constructed. Based on the convex hull boundary extraction results, the external morphological features of the data points are determined, including the boundary point sequence and the convex hull polygon structure.

3. The method for detecting icing on transmission lines according to claim 2, characterized in that, The preprocessed signal is converted into a set of discrete 3D data points using a 3D coordinate mapping function, where each data point contains spatial coordinate information, including: The preprocessed signal is analyzed in the time domain to obtain the timestamp, amplitude and frequency parameters of each sampling point in the signal; Based on the transformation relationship between the radar coordinate system and the spatial world coordinate system, the timestamp, amplitude and frequency parameters of each sampling point are used as input, and their X-axis, Y-axis and Z-axis coordinate values ​​in three-dimensional space are calculated respectively through the three-dimensional coordinate mapping function; The spatial coordinates calculated from all sampling points are aggregated to generate a three-dimensional discrete data point set, where each data point is uniquely identified by its corresponding spatial coordinates.

4. The method for detecting icing on transmission lines according to claim 3, characterized in that, The discrete data point set is sorted by polar angle, and starting from the lowest point, all data points are traversed using the Graham scan algorithm. The convexity of the point set is determined by the cross product of vectors, and the convex hull boundary is constructed step by step, including: In the set of discrete three-dimensional data points, the point with the smallest Z coordinate value is selected as the origin of the coordinate system. If there are multiple points with the same smallest Z value, the point with the smallest Y coordinate value is selected. If there are still multiple points with the same smallest Y value, the point with the smallest X coordinate value is selected as the final starting point. Using the starting point as the pole, calculate the polar angles of all other data points relative to this pole, and sort them in order of increasing polar angle. The data points are sorted in a large order to form a sorted sequence of data points; Initialize an empty stack, push the starting point and the first two data points in the sorted sequence onto the stack in sequence, and then traverse the remaining data points in sorted order. For each current traversal point, calculate the cross product of the two vectors formed by the second point on the top of the stack, the stack vertex, and the current traversal point. Based on the cross product result, determine whether the current traversal point causes the stack vertex to be a non-convex hull vertex. If so, pop the stack vertex. Continue until the stack top satisfies the convexity condition, then push the current traversal point onto the stack. In this way, the convex hull boundary is gradually constructed. After the traversal is complete, the sequence of points stored in the stack is the final extracted convex hull boundary data points.

5. The method for detecting icing on transmission lines according to claim 4, characterized in that, Based on the convex hull boundary extraction results, the peripheral morphological features of the data points are determined, including the boundary point sequence and the convex hull polygon structure, including: Read the sequence of convex hull boundary data points, and connect the data points sequentially according to their storage order in the stack to form an initial boundary polygon; The initial boundary polygon is subjected to vertex optimization processing, which removes collinear redundant vertices, retains key turning vertices, and generates an optimized boundary point sequence. Based on the optimized boundary point sequence, calculate its geometric center point, and use the center point as a reference to emit rays to each boundary point, connecting the boundary point sequence to form a closed convex hull polygon structure. The boundary point sequence and convex hull polygon structure are output as the peripheral morphological features of the data points.

6. The method for detecting icing on transmission lines according to claim 5, characterized in that, The clean transmission signal is input into a pre-trained deep convolutional neural network, and the spatial properties of the ice layer are extracted through multiple convolutional and pooling layers. Feature fusion is performed after a fully connected layer to calculate the ice layer thickness distribution, including: The clean transmission signal is converted into a three-dimensional feature tensor; Based on the three-dimensional feature tensor, primary spatial features are extracted through the first convolutional layer, and the extracted primary spatial features are reduced in dimensionality using the first pooling layer to obtain a primary feature map containing key feature information. Based on the primary feature map, intermediate spatial features are extracted through a second convolutional layer, and the intermediate spatial features are compressed using a second pooling layer to generate an intermediate feature map with enhanced expressive power. Based on the intermediate feature map, high-level spatial features are extracted through the third convolutional layer to capture detailed information about the ice layer distribution and generate a high-level feature map. The high-level feature map is flattened into a one-dimensional feature vector and input into a fully connected layer for feature fusion and transformation to obtain the fused features. Based on the fused features, the ice thickness value corresponding to each spatial location is calculated through the regression output layer to generate an ice thickness distribution map.

7. The method for detecting icing on transmission lines according to claim 6, characterized in that, Ice layer categories are identified using a softmax classifier. Based on the thickness distribution map and the ice layer category identification results, detection results containing ice layer thickness values ​​and spatial distribution attributes are generated, including: Based on the fused features output by the fully connected layer, a feature vector for classification is extracted and input into the softmax classifier. The softmax classifier processes the input feature vector to calculate the probability distribution of each ice layer category, and selects the category with the highest probability as the final identification result of the ice layer category. Based on the ice thickness distribution map generated by the regression output layer and the ice category identification result obtained by the softmax classifier, spatial alignment and feature fusion processing are performed to obtain fused data containing thickness and category information. Based on the fused data, a detection result is generated that includes ice layer thickness values, spatial distribution attributes, and category information.

8. A transmission line icing detection system, characterized in that, The system is used to perform the method as described in any one of claims 1 to 7, comprising: The acquisition module is used to acquire raw echo data from the surface of the transmission line and to preprocess the raw echo data to obtain a preprocessed signal. The correction module is used to convert the preprocessed signal into a set of discrete 3D data points, extract the convex hull boundary of the discrete 3D data point set based on the Graham scan algorithm, and determine the outer morphological features of the data points; establish multiple reference ring templates within the outer morphological features using a midpoint closed curve generation algorithm, divide the region according to the density distribution characteristics of the data points, and generate multiple feature analysis subdomains; calculate the spatial compensation coefficients using principal component analysis based on the curvature attributes and spatial arrangement of the data points in each feature analysis subdomain, and use the spatial compensation coefficients to correct the position parameters of the data points to obtain an optimized processed signal; The elimination module is used to eliminate transmission interference and obtain a clean transmission signal during channel switching by using the Manchester coding protocol and local buffer verification algorithm. The prediction module is used to input the clean transmission signal into a pre-trained deep convolutional neural network, extract the spatial attributes of the ice layer through multiple convolutional and pooling layers, perform feature fusion through a fully connected layer, calculate the ice layer thickness distribution, identify the ice layer category through a softmax classifier, and generate a detection result containing the ice layer thickness value and spatial distribution attributes based on the thickness distribution map and the ice layer category identification result.

Images