A method and system for ice thickness measurement based on point cloud

Abstract

The application discloses an ice-coating thickness measurement method and system based on a point cloud, and relates to the field of transmission line online monitoring. The conductor point cloud and the tower point cloud are stably extracted from the complex line corridor point cloud through an arc-drop perception point cloud segmentation network, the single-conductor point set is separated through span division and geometric clustering, the three-dimensional center line of the conductor is reconstructed through weighted fitting, the current arc drop is calculated through the hanging point identification and the arc drop definition, the arc-drop increment is obtained by comparison with the historical ice-free point cloud, and the equivalent horizontal tension is inversely calculated, finally, the equivalent ice-coating thickness is calculated based on the circular ring equivalent model through the inverse calculation of the additional load based on the arc-drop increment, and the engineering problem that the outer contour of the ice-coated conductor is difficult to stably obtain under the condition of the unmanned aerial vehicle laser point cloud is solved, so that the ice-coating state is effectively evaluated.

Description

Technical Field

[0001] This application relates to the field of online monitoring of power transmission lines, and in particular to a method and system for measuring ice thickness based on point clouds. Background Technology

[0002] Overhead power transmission and distribution lines are prone to conductor icing under low temperatures, rain, snow, freezing rain, and microclimate conditions in mountainous areas. Icing increases the load per unit length of the conductor, causing increased conductor sag, reduced distance to ground, and potentially leading to line breaks, tower collapses, phase-to-phase flashovers, and large-scale power outages. Therefore, accurately determining the icing status of conductors is a crucial foundation for icing prevention, de-icing, and disaster early warning systems for power transmission and distribution lines.

[0003] Existing methods for monitoring icing mainly include manual ice observation, fixed icing monitoring terminals, image recognition thickness measurement, tensile sensor inversion thickness measurement, and direct thickness measurement using laser point clouds. Manual ice observation is greatly affected by terrain and meteorological conditions; fixed monitoring terminals have limited coverage; image recognition methods are easily affected by lighting, fog, obstruction, and background interference; direct thickness measurement using laser point clouds usually requires a relatively complete acquisition of the outer contour of the iced conductor, but under UAV mobile acquisition conditions, the outer contour of the icing is often difficult to reconstruct stably due to the influence of flight attitude, scanning angle, point cloud density, and ice reflectivity.

[0004] On the other hand, the essential impact of icing on conductors is to increase the load per unit length and cause changes in the conductor's spatial morphology, with increased sag being a significant manifestation. Methods for calculating equivalent icing thickness based on sag changes already exist, but these methods often rely on fixed measuring devices or specific measuring points, making them unsuitable for multi-span, non-contact observations under UAV-based mobile laser point cloud conditions. While UAV laser point clouds can cover relatively long lines, the conductor targets are slender, the point cloud is sparse with significant local gaps, and it is easily affected by interference from towers, insulator strings, vegetation, and background points. If only conventional point cloud semantic segmentation methods are used, only conductor category labels are often obtained, making it difficult to reliably obtain the conductor's three-dimensional centerline, suspension point position, and lowest point position for sag calculation.

[0005] Under the conditions of mobile data collection by UAVs, the laser point cloud of transmission lines has problems such as thin and sparse conductor targets, severe local missing parts, and susceptibility to interference from towers, vegetation and background points. As a result, it is difficult to obtain the outer contour of the icing conductor stably. Traditional methods for measuring ice thickness based on direct measurement of the outer contour or relying on fixed devices have low accuracy and poor adaptability, and cannot effectively assess the icing status.

[0006] Therefore, there is an urgent need for a UAV laser point cloud processing method for icing sag inversion tasks, which can not only extract the conductor point cloud, but also output the conductor centerline, centerline confidence and geometric parameters required for sag calculation, and further combine it with the icing equivalent thickness inversion model to realize mobile, non-contact, span-level icing state calculation. Summary of the Invention

[0007] To address the aforementioned technical problems, this application provides a point cloud-based method for measuring icing thickness, comprising: Acquire UAV laser point cloud data of the transmission line under test in its current state, and preprocess the UAV laser point cloud data to extract the initial point features of each point. The initial point features are input into the sag-sensing point cloud segmentation network for point-level semantic segmentation, outputting point-level semantic results for conductors, towers, and background. Based on the point-level semantic results, conductor point clouds and tower point clouds are extracted from the UAV laser point cloud data. The sag-sensing point cloud segmentation network includes a structure prior embedding module, an encoding module, a point-level decoding module, and a semantic prediction head, which are used to enhance the ability to distinguish between the slender continuous structure of conductors and the vertical structure of towers. Based on the tower point cloud and the tower spatial location, the span range between adjacent towers is determined, and within each span range, candidate conductor point clouds are extracted from the conductor point cloud. Single conductor separation is performed on the candidate conductor point cloud to obtain a single conductor point set. Based on the conductor category probability, local linearity, local point cloud density, and fitting residual confidence of each conductor point in the single conductor point set, the point-level fitting confidence of each conductor point is calculated, and a weighted catenary fitting is performed on the single conductor point set based on the point-level fitting confidence to obtain the three-dimensional centerline of the corresponding conductor. Identify the hanging points at both ends of the conductor based on the spatial adjacency relationship between the two ends of the three-dimensional centerline and the corresponding tower point cloud, and use the connecting line of the two hanging points as the reference chord to calculate the maximum vertical difference between the three-dimensional centerline and the reference chord to obtain the current sag. The sag increment is determined based on the current sag and the reference sag. The equivalent horizontal tension is calculated based on the reference sag, span length, and conductor foundation information under historical ice-free conditions. The additional load per unit length caused by icing is inverted based on the sag increment, span length, and equivalent horizontal tension. The unit length additional load is converted into unit length icing mass, and the equivalent icing thickness of the transmission line under test is calculated based on the bare conductor radius, icing density, and the circular equivalent model.

[0008] In one embodiment, the sag-aware point cloud segmentation network includes a structure prior embedding module, an encoding module, a point-level decoding module, and a semantic prediction head. The structure prior embedding module is used to describe the linearity, vertical distribution features and local density features of each point cloud in the local neighborhood, construct the structure prior features of each point cloud, and obtain the structure embedding features. The encoding module includes several encoding layers for performing layer-by-layer encoding processing on the structural embedding features. Each encoding layer includes a local geometric encoding branch, a directional continuous encoding branch, and a structure-gated fusion unit. The local geometric encoding branch is used to extract spatial morphological differences in the neighborhood to obtain local geometric features. The directional continuous encoding branch is used to extract strip-level continuous features along the span direction to obtain directional continuous enhanced features. The structure-gated fusion unit is used to adaptively weight and fuse the local geometric features and the directional continuous enhanced features to output the fused point features of each layer. The point-level decoding module is used to perform point-level semantic decoding on the fused point features of each layer, and to fuse the fused point features of each layer to obtain the decoded point-level features. The semantic prediction head is used to output the predicted probability of each point belonging to the conductor category or the tower category based on the decoded point-level features, and take the category corresponding to the highest probability as the semantic label of each point to generate the point-level semantic result.

[0009] In one embodiment, the step of obtaining structural embedding features through a structural prior embedding module includes: Based on the three-dimensional coordinates, laser reflection intensity, and neighborhood point cloud distribution of each point, the local point cloud density, local linear structure response, and vertical structure response of each point are calculated. The three-dimensional coordinates, laser reflection intensity, local point cloud density, local linear structure response, and vertical structure response are stitched together to obtain the enhanced input features; The enhanced input features are upscaled using a multilayer perceptron and mapped to a preset feature channel dimension to obtain the structural embedding features.

[0010] In one embodiment, the processing steps of the encoding module include: Through the local geometric coding branch, the farthest point sampling is performed on the point cloud corresponding to the structural embedding features of the current layer to obtain a low-resolution point set; A preset number of nearest neighbor sets are established for each sampling point in the low-resolution point set, and the local geometric features are output by combining the local geometric feature mapping function and the neighborhood max pooling operation. By using the directional continuous coding branches, the local span direction is determined according to the spatial location of the tower or the route, and the projected coordinates of each sampling point along the local span direction are calculated. The laser point cloud data is divided into directional strips according to the projection coordinates, the point features in each directional strip are aggregated by pooling to obtain strip-level features, and one-dimensional continuous feature propagation is performed on the strip-level features of adjacent strips along the arrangement direction of the strips, and the propagated strip-level features are fed back to the points in the corresponding strips to obtain the directionally continuous enhanced features; The local geometric features and the directionally continuous enhanced features are spliced along the feature channel dimension through the structure-gated fusion unit to obtain spliced features; The spliced features are input into a gated multi-layer perception machine and mapped to a preset interval combined with an activation function to determine a gating weight; The local geometric features and the directionally continuous enhanced features are weighted by the gating weight, and the two weighted results are added element by element to obtain the fused point features of each layer.

[0011] In an embodiment, the step of obtaining the decoded point-level features through the point-level decoding module includes: The input features of the current decoding layer are obtained, and the input features of the current decoding layer are processed by nearest neighbor interpolation to obtain interpolated features, wherein the number of feature points of the interpolated features is restored to be consistent with the number of fused point features corresponding to the current decoding layer; The interpolated features and the fused point features output by the corresponding encoding stage are spliced along the feature channel dimension, and feature fusion is performed through a decoding multi-layer perception machine to obtain the point-level features.

[0012] In an embodiment, the step of extracting a conductor candidate point cloud from the conductor point cloud in each range and separating the conductor candidate point cloud into single conductors to obtain a single conductor point set includes: In the conductor candidate point cloud in each range, a spatial adjacency relationship is established according to the comparison result of the three-dimensional spatial distance between each point and a preset neighborhood radius threshold; Based on the spatial adjacency relationship as the connection basis, a connected clustering is performed on all conductor candidate points to determine an initial clustering cluster; The principal direction of the initial clustering cluster is calculated by principal component analysis, and the point cloud is divided into at least two slices along the principal direction, and the distribution peaks of the point cloud in the horizontal and vertical directions of each slice are counted; The point cloud at the corresponding slice position is segmented into a preset number of sub-clusters along the gap between the distribution peaks, and the segmentation operation is iteratively performed until the horizontal and vertical distributions of each sub-cluster in the principal direction are single peaks, and the final sub-cluster is determined as a single conductor point set corresponding to a spatially independent continuous conductor.

[0013] In an embodiment, the step of calculating the point-level fitting confidence of each wire point according to the wire class probability, local linearity, local point cloud density and fitting residual confidence of each wire point in the single-wire point set, and performing weighted catenary fitting on the single-wire point set based on the point-level fitting confidence to obtain the three-dimensional center line of the corresponding wire further comprises: obtaining the wire class probability, local linearity and local point cloud density of each wire point in the single-wire point set; calculating the fitting residual confidence according to the distance of the wire point to the initial fitting curve and the scale parameter; weighting calculating the point-level fitting confidence of each wire point according to the wire class probability, local linearity, local point cloud density and fitting residual confidence; performing random sampling and initial curve fitting on the single-wire point set using the random sample consensus algorithm to obtain an inner point set: In a local line route coordinate system, using a catenary model to describe the vertical coordinates of the wire along the span direction, using a transverse model to describe the coordinates of the wire along the transverse direction, weighting least squares fitting on the inner point set with the point-level fitting confidence as the weight, and solving the catenary model parameters and the transverse model parameters to obtain the three-dimensional center line of the wire.

[0014] In an embodiment, the step of identifying the wire end hanging points according to the spatial adjacency relationship between the two ends of the three-dimensional center line and the corresponding tower point cloud, and taking the connecting line of the two end hanging points as the reference chord line to calculate the maximum vertical difference of the three-dimensional center line relative to the reference chord line to obtain the current sag comprises: determining the spatial coordinates of the two end hanging points, and taking the connecting line of the two end hanging points as the reference chord line; calculating the vertical coordinates of the reference chord line at each span direction position; obtaining the vertical coordinates of the three-dimensional center line at the same span direction position, calculating the difference between the vertical coordinates of the reference chord line and the vertical coordinates of the three-dimensional center line, and taking the maximum value of the difference between the two end hanging points as the current sag.

[0015] The step of converting the unit length additional load into unit length ice mass, and calculating the equivalent ice thickness of the to-be-measured power transmission line based on the wire bare radius, ice density and circular ring equivalent model further comprises: calculating the result confidence of the ice thickness measurement result according to the wire point cloud integrity, wire segmentation quality, single-wire clustering continuity, three-dimensional center line fitting residual, hanging point identification quality and historical ice-free benchmark matching quality; when the result confidence is lower than a preset threshold, outputting a low confidence label; When the wire hanging point is missing, the three-dimensional center line fitting fails, or the historical ice-free benchmark cannot be matched, an uncalculated label is output.

[0016] In addition, to achieve the above-mentioned purpose, the application also provides a point cloud-based ice thickness measurement system for executing the point cloud-based ice thickness measurement method as described above; the point cloud acquisition and preprocessing module, the sag-aware point cloud segmentation module, the span division module, the single wire separation module, the three-dimensional center line fitting module, the hanging point identification and sag calculation module, and the ice thickness inversion module; The point cloud acquisition and preprocessing module is configured to acquire unmanned aerial vehicle laser point cloud data of a power transmission line to be measured in a current state, and pre-process the unmanned aerial vehicle laser point cloud data to extract initial point features. The sag-aware point cloud segmentation module is configured to perform point-level semantic segmentation on the initial point features based on a sag-aware point cloud segmentation network to obtain point-level semantic results of wires, towers and backgrounds, and extract wire point clouds and tower point clouds. The span division module is configured to determine a span range between adjacent towers according to the tower point clouds and spatial positions of the towers. The single wire separation module is configured to extract wire candidate point clouds from the wire point clouds in each span range, and separate the wire candidate point clouds to obtain a single wire point set. The three-dimensional center line fitting module is configured to calculate a point-level fitting confidence based on a wire category probability, a local linearity, a local point cloud density and a fitting residual confidence, and perform weighted catenary fitting on the single wire point set based on the point-level fitting confidence to obtain a three-dimensional center line of a corresponding wire. The hanging point identification and sag calculation module is configured to identify two end hanging points of a wire according to a spatial adjacency relationship between two ends of the three-dimensional center line and corresponding tower point clouds, and calculate a maximum vertical difference of the three-dimensional center line relative to a reference chord line formed by connecting the two end hanging points to obtain a current sag. The ice thickness inversion module is configured to invert a unit length additional load caused by ice based on the current sag, a benchmark sag, a span length and wire basic information in a historical ice-free state; convert the unit length additional load into a unit length ice mass, and calculate an equivalent ice thickness of the power transmission line to be measured based on a wire bare radius, an ice density and a circular ring equivalent model.

[0017] In addition, the point cloud-based ice thickness measurement method and system provided by the application at least have the following technical effects: First, the dependence on the integrity of the ice layer outer contour is reduced.

[0018] Traditional point cloud thickness measurement method usually needs to obtain the outer surface of the icing conductor completely, and estimate the ice thickness by the change of the outer diameter. When the point cloud density is insufficient, the scanning angle is limited, or the ice layer distribution is uneven, the outer diameter measurement is easy to produce larger error. The present application inverses the icing load by the conductor sag increment, and then calculates the equivalent icing thickness, without requiring complete recovery of the ice layer outer contour, and is suitable for icing measurement under the condition of unmanned aerial vehicle mobile scanning.

[0019] Secondly, the stability of conductor extraction in complex line corridor is improved.

[0020] The SAG-Net of the present application introduces the conductor linear structure, the vertical structure of the tower and the direction continuity feature, which can better distinguish the conductor points, the tower points and the vegetation / background points. The design can reduce the vegetation misclassification and the conductor breakage problem, and provide a more reliable conductor point cloud basis for subsequent catenary fitting.

[0021] Thirdly, the sag calculation is more in line with the actual shape of the transmission conductor.

[0022] The transmission conductor is in a sag curve shape under the action of gravity and tension. Compared with ordinary polynomial fitting or local slice fitting, the present application uses a catenary model to reconstruct the conductor center line, which is more in line with the actual stress shape of the conductor. Combined with RANSAC denoising and point-level fitting confidence weighting mechanism, the influence of misclassified points and outliers on sag calculation can be reduced.

[0023] Fourthly, the reference sag is calculated from historical ice-free point cloud, reducing the error of pure theoretical calculation.

[0024] The present application calculates the reference sag from the historical ice-free point cloud of the same line, instead of relying entirely on the theoretical design sag. Since the historical ice-free point cloud can reflect the real running state of the line, when compared with the current icing point cloud, the influence of construction error, long-term running state change and theoretical parameter deviation on icing measurement can be reduced.

[0025] Fifthly, the result confidence can be output, reducing the risk of misjudgment.

[0026] The present application includes the conductor segmentation quality, single conductor clustering continuity, catenary fitting residual, hanging point recognition quality and historical ice-free reference matching quality into the result confidence evaluation. When the conductor point cloud is seriously missing, the hanging point cannot be reliably identified, or the center line fitting residual is too large, low confidence or uncalculated label can be output, avoiding forced output of the icing thickness result when the data quality is insufficient. BRIEF DESCRIPTION OF DRAWINGS

[0027] Figure 1 The flowchart of the embodiment of the point cloud-based icing thickness measurement method of the present application; Figure 2A schematic diagram of an overall flow of an ice thickness measurement method based on point clouds according to an embodiment of the present application; Figure 3 A schematic diagram of a SAG-Net network structure involved in an embodiment of the present application; Figure 4 A schematic diagram of a feature extraction module structure involved in an embodiment of the present application; Figure 5 A schematic diagram of a UAV laser radar collection and span division involved in an embodiment of the present application; Figure 6 A schematic diagram of a conductor center line reconstruction and sag calculation involved in an embodiment of the present application; Figure 7 A schematic diagram of a hardware running environment architecture of a terminal device involved in an embodiment of the present application. DETAILED DESCRIPTION

[0028] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the scope of protection of the present application.

[0029] It should be noted that if the embodiments of the present application involve directional indications, such as up, down, left, right, front, back, etc., the directional indications are only used to explain the relative position relationship, movement, etc. between components in a certain posture, and if the certain posture changes, the directional indications also change accordingly. In addition, if the embodiments of the present application involve descriptions such as “first, second”, “S1, S2”, “step one, step two” and the like, such descriptions are only for description purposes, and cannot be understood as indicating or implying the relative importance of the technical features indicated or the number of technical features indicated, or indicating the execution order of the method, etc. Those skilled in the art can understand that any change within the technical concept of the invention without deviating from the invention points should be included in the protection scope of the present application.

[0030] As shown in Figure 1 The flowchart of an embodiment of an ice thickness measurement method based on point clouds according to the present application is shown in Figure 1 A schematic diagram of an overall flow of an ice thickness measurement method based on point clouds according to an embodiment of the present application, including: Figure 2 An embodiment of an ice thickness measurement method based on point clouds according to the present application is given below, Figure 2 A schematic diagram of an overall flow of an ice thickness measurement method based on point clouds according to an embodiment of the present application, including: Step S10: Obtain UAV laser point cloud data of a power transmission line to be measured in a current state, and pre-process the UAV laser point cloud data to extract initial point features of each point.

[0031] In the embodiment, the laser point cloud data is a set of three-dimensional space points collected by an unmanned aerial vehicle (UAV) equipped with an airborne laser radar along the power transmission line during round-trip flight, and each point contains its three-dimensional coordinate information in space and laser echo intensity information. Preprocessing refers to a series of cleaning and standardization operations performed on the original point cloud data, including denoising, downsampling, and coordinate unification. Among them, denoising is used to remove outliers caused by atmospheric scattering, multipath effect, or sensor noise; downsampling is used to control the overall density of the point cloud while preserving the slender structure characteristics of the conductor, avoiding excessive data volume affecting processing efficiency; coordinate unification is used to register multiple frames of point cloud collected by the UAV during round-trip flight to the same spatial coordinate system. The local line coordinate system is a local spatial reference frame established based on the direction of the power transmission line, with the horizontal axis extending along the span direction, the vertical axis being the transverse direction perpendicular to the line direction, and the vertical axis being the vertical direction. The target feature dimensionality is to map the original input features of each point to a higher-dimensional feature space through a multi-layer perception machine, so that the network can learn the complex relationship between points in a higher-dimensional representation.

[0032] Specifically, the processing action is performed by the ice thickness measurement system. First, the laser point cloud data collected by the UAV onboard laser radar along the power transmission line during round-trip flight is obtained. Statistical filtering operation is performed on the laser point cloud data: for each point in the laser point cloud data, the average distance of all points in its neighborhood to the point is calculated, and if the average distance exceeds the sum of the global average distance and the preset standard deviation multiple, the point is determined as an outlier and is removed, thereby eliminating isolated noise points caused by atmospheric scattering or sensor noise.

[0033] Then, the laser point cloud data is downsampled using a voxel downsampling method: the three-dimensional space is divided into a voxel grid with a preset edge length, and the center of gravity coordinates of all points falling into the same voxel grid are used to replace the original point cloud, thereby effectively controlling the background point cloud density while preserving the slender structure characteristics of the conductor, avoiding excessive data volume affecting the subsequent processing efficiency.

[0034] Then, the multi-frame point clouds collected by the unmanned aerial vehicle through round-trip flight are registered to the same spatial coordinate system by an iterative closest point algorithm or a feature matching method based on significant landmarks such as towers, so as to eliminate the coordinate deviation between the multi-frame data. After the above preprocessing is completed, a local line coordinate system is established according to the overall trend of the power transmission line and the spatial position of the tower: taking the line trend direction as the horizontal axis direction, taking the horizontal direction perpendicular to the horizontal axis as the vertical axis direction, and taking the vertical direction as the vertical axis direction, the spatial coordinates of all points are converted from the original coordinate system to the local line coordinate system. Finally, the three-dimensional coordinates, laser reflection intensity, local point cloud density, local linear structure response and vertical structure response of each point after preprocessing are extracted as target features, and these target features are spliced into enhanced input feature vectors, and the enhanced input feature vectors are mapped from the original feature dimension to the preset high-dimensional feature space through a multi-layer perception machine, to obtain the initial point features corresponding to each point.

[0035] As an optional embodiment of preprocessing the laser point cloud data, the laser point cloud data collected by the unmanned aerial vehicle on-board laser radar is inputted: ;

[0036] Among them, represents the batch size, represents the number of input point cloud points, represents the number of input feature channels of each point. The basic features of each point at least include three-dimensional coordinate information: ;

[0037] Among them, i represents the serial number (index) of the point in the point cloud, which is used to distinguish different points, represents the coordinate value of the i th point along the horizontal axis (span direction) in the local line coordinate system (or the original spatial coordinate system), represents the coordinate value of the i th point along the vertical axis (transverse direction perpendicular to the line trend) in the local line coordinate system, represents the coordinate value of the i th point along the vertical axis (vertical direction, i.e. elevation direction) in the local line coordinate system; in the case of having laser reflection intensity, echo information or local geometric features, the input features of each point can be extended to: ;

[0038] Among them, represents the laser reflection intensity, represents the local density of the point cloud, represents the local linearity.

[0039] Before inputting into the network, the point cloud data is processed for coordinate unification, denoising, downsampling and normalization. Specifically, it includes: 1) Unify the multi-frame point clouds collected by the unmanned aerial vehicle in round trips to the same spatial coordinate system; 2) Remove outliers by statistical filtering or radius filtering; 3) Control the background point cloud density by voxel downsampling while trying to preserve the conductor slender structure as much as possible; 4) Establish a local line coordinate system according to the line direction and the spatial position of the tower, wherein the axis is in the span direction, the axis is in the transverse direction, and the axis is in the vertical direction.

[0040] Subsequently, the input point features are upgraded by a multi-layer perception (MLP) to obtain initial point features:

[0041] wherein, represents the feature dimension after upgrading.

[0042] By performing denoising, downsampling and coordinate unification on the original point cloud, the influence of sensor noise and flight attitude difference on data quality is eliminated, providing clean and standardized input data for subsequent semantic segmentation; by establishing a local line coordinate system, the distribution rule of the conductor point cloud in the span direction is explicitly expressed; by upgrading the target features, the low-dimensional spatial information is mapped to a high-dimensional feature space, enhancing the network's feature expression ability for the conductor slender structure and the tower vertical structure, and laying a data foundation for high-precision semantic segmentation of the sag-aware point cloud segmentation network.

[0043] Step S20: inputting the initial point features into the sag-aware point cloud segmentation network for point-level semantic segmentation, outputting point-level semantic results of the conductor, the tower and the background, and extracting the conductor point cloud and the tower point cloud from the unmanned aerial vehicle laser point cloud data according to the point-level semantic results; wherein the sag-aware point cloud segmentation network comprises a structure prior embedding module, an encoding module, a point-level decoding module and a semantic prediction head, for enhancing the distinguishing ability of the conductor slender continuous structure and the tower vertical structure.

[0044] In this embodiment, the sag-aware point cloud segmentation network is a deep learning semantic segmentation network designed for the transmission line sag calculation task, and its core function is to identify and distinguish conductor points and tower points from complex line corridor point clouds, and it can also identify and distinguish background points. Semantic segmentation is a process of assigning a class label to each point in the point cloud, and the class label includes but is not limited to three class labels of conductor, tower and background.

[0045] ​As an optional implementation, the application adopts the sagging perception point cloud segmentation network SAG-Net to extract the conductor point cloud, the tower point cloud and the vegetation / background point cloud. The network can also be replaced by the existing point cloud semantic segmentation model such as PointNet++, KPConv, RandLA-Net, Point Transformer, PointNeXt or sparse convolution network. The above-mentioned model can also realize the semantic separation between the conductor, the tower and the background.

[0046] However, the general point cloud segmentation network is usually not specially designed for the elongation, sparsity and continuity along the span direction of the power conductor, and under the condition of complex vegetation background or local missing of the conductor point cloud, the conductor is prone to be broken, misclassified and missed. The application preferably adopts SAG-Net, which improves the continuity and stability of the conductor point cloud extraction through structure prior embedding, direction continuity coding and structure gating fusion mechanism.

[0047] The sagging perception point cloud segmentation network is composed of a structure prior embedding module, an encoding module and a lightweight point-level decoding module. The structure prior embedding module is used to encode the geometric structure prior knowledge of the point cloud such as the elongated linear structure of the conductor and the vertical lattice structure of the tower into a network learnable feature representation, so that the network can utilize these prior knowledge to enhance the recognition ability of the target structure during training and inference. The encoding module includes a local geometric coding branch, a direction continuity coding branch and a structure gating fusion unit: the local geometric coding branch captures local geometric features by constructing a point cloud neighborhood graph and extracting spatial geometric differences within the neighborhood; the direction continuity coding branch enhances the continuity expression of the conductor along the strike direction by dividing the point cloud into strips along the span direction and propagating the continuity features between the strips, thereby outputting direction continuity enhanced features. The structure gating fusion unit is used to adaptively fuse the local geometric features and the direction continuity enhanced features, dynamically adjusts the weight distribution of the two types of features according to the characteristics of the local structure of each point, and outputs the fused point features. The point-level decoding module can be a lightweight point-level decoding module, which is used to restore the low-resolution features after multiple downsampling and feature abstraction to the original point cloud resolution layer by layer, and outputs the prediction probability of the class of each point through a semantic prediction head.

[0048] Specifically, the initial point features are input into the sagging perception point cloud segmentation network SAG-Net to perform semantic recognition on the conductor, tower and vegetation / background in the power line point cloud. The SAG-Net is oriented to the sagging calculation task of the power line, and through the conductor elongated structure prior, the tower vertical structure prior and the direction continuity constraint, the separability of the conductor point cloud in the complex background is enhanced, and the semantic prediction result of each point is output: ;

[0049] wherein, denotes batch size, denotes input points, corresponding to three semantic labels of conductor, tower, and vegetation / background, i.e.: ;

[0050] wherein, denotes conductor points, denotes tower points, denotes vegetation, ground, building, and other background points.

[0051] The SAG-Net includes a structure prior embedding module, a local geometry encoding branch, a direction continuous encoding branch, a structure gating fusion unit, and a lightweight point-level decoding module. The network does not directly output the span, the hanging point, or the conductor centerline, but provides reliable conductor point clouds and tower point clouds for subsequent span division, hanging point identification, and catenary fitting.

[0052] This step realizes high-precision semantic separation of the conductor point cloud, the tower point cloud, and the background point cloud through the sag-aware point cloud segmentation network. The structure prior embedding module enables the network to utilize the slender linear structure of the conductor and the vertical lattice structure of the tower as prior knowledge, thereby enhancing the network's recognition ability of the target structure. The direction continuous encoding branch effectively solves the problem of conductor point cloud breakage caused by local occlusion or point cloud sparseness through strip division and continuous feature propagation along the span direction, so that the segmented conductor point cloud is more continuous and complete. The structure gating fusion unit adaptively adjusts the weight distribution of the local geometric features and the direction continuous features according to the local structure characteristics of each point, so that the conductor points obtain stronger direction continuity constraints, the tower points and the background points obtain stronger local geometric constraints, thereby forming a more stable structure distinction between the conductor, the tower, and the background. The above designs collectively ensure that high-quality, continuous, and complete conductor point clouds and tower point clouds are extracted from complex line corridor point clouds, thereby providing a reliable data basis for subsequent span division, single conductor separation, and sag calculation.

[0053] Optionally, in this embodiment, the sag-aware point cloud segmentation network includes a structure prior embedding module, an encoding module, a point-level decoding module, and a semantic prediction head: The structure prior embedding module is configured to describe the linear degree, the vertical distribution feature, and the local density feature of each point cloud in a local neighborhood, construct the structure prior feature of each point cloud, and obtain structure embedding features. Step S22: an encoding module, comprising a plurality of encoding layers, is configured to perform layer-by-layer encoding processing on the structure embedding feature; each encoding layer comprises a local geometry encoding branch, a direction continuity encoding branch, and a structure gate fusion unit; the local geometry encoding branch is configured to extract spatial form differences of a neighborhood to obtain a local geometry feature; the direction continuity encoding branch is configured to extract a strip-level continuity feature along a range direction to obtain a direction continuity enhanced feature; and the structure gate fusion unit is configured to perform adaptive weighted fusion on the local geometry feature and the direction continuity enhanced feature to output a fused point feature of each layer; Step S23: a point-level decoding module is configured to perform point-level semantic decoding on the fused point feature of each layer, and fuse the fused point features of the layers to obtain a decoded point-level feature. Step S24: a semantic prediction head is configured to output a prediction probability of each point belonging to a wire category or a tower category according to the decoded point-level feature, take a category corresponding to a maximum probability as a semantic label of each point, and generate the point-level semantic result.

[0054] Specifically, the initial point feature of each point obtained in step S10 is input into the sagitta-aware point cloud segmentation network. First, the initial point feature is enhanced by a structure prior embedding module: for each point in the point cloud, a local linear structure response is calculated based on three eigenvalues of a neighborhood covariance matrix of the point, and a vertical structure response is calculated based on an included angle between a neighborhood principal direction and a vertical direction; the three-dimensional coordinates, the laser reflection intensity, the local point cloud density, the local linear structure response, and the vertical structure response are spliced into enhanced input features; and a multi-layer perceptron is used to map the enhanced input features from an original dimension to a preset high-dimensional feature space to output a structure embedding feature.

[0055] Then, the structural embedding features are encoded by an encoding module: in the local geometry encoding branch, farthest point sampling is performed on the point cloud of the current layer to obtain a low-resolution point set and its corresponding low-resolution features, a preset number of nearest neighbor sets are established for each sampling point in the original point cloud, and the features, relative coordinate offsets, offset distances, local linear structure responses, and vertical structure responses of the neighboring points are input into a local geometry feature mapping multi-layer perception, and the local geometry features are output after neighborhood max-pooling; in the direction continuous encoding branch, the local span direction is determined according to the tower space position or the line direction, the projection coordinates of the points along the direction are calculated, the point cloud of the current layer is divided into multiple direction strips according to the projection coordinates, the point features in each direction strip are pooled and aggregated to obtain strip-level features, and the strip-level features of adjacent strips are propagated along the strip arrangement direction to obtain direction continuous enhanced features. The local geometry features and the direction continuous enhanced features are adaptively fused by a structure gating fusion unit: the local geometry features and the direction continuous enhanced features corresponding to the same sampling point are spliced along the feature channel dimension, the spliced features are input into a gated multi-layer perception, and the gating weights are obtained by mapping the Sigmoid activation function to the interval [0, 1] to obtain the gating weights. The local geometry features are weighted by the gating weights, and the direction continuous enhanced features are weighted by one minus the gating weights. The two weighted results are added element by element to obtain the fused point features.

[0056] Finally, the fused point features are restored to the original point cloud resolution by a point-level decoding module: for the input features of the current decoding layer, the number of feature points is restored to the number of fused point features output in the encoding stage corresponding to the decoding layer by nearest neighbor interpolation or inverse distance weighted interpolation, the interpolated features are spliced with the fused point features output in the corresponding encoding stage along the feature channel dimension, and the spliced features are fused by a decoding multi-layer perception. The above interpolation and splicing fusion operations are repeatedly performed until the number of feature points is restored to the number of the original input point cloud, and the original resolution point-level features are obtained. The original resolution point-level features are input into a classification multi-layer perception and normalized by a Softmax function, and the prediction probability of each point belonging to the conductor category or the tower category is output. The category corresponding to the maximum probability is taken as the semantic label of each point according to the prediction probability, and a semantic prediction result is generated. All points labeled as conductor points are extracted from the point cloud as conductor point cloud, and all points labeled as tower points are extracted as tower point cloud.

[0057] Specifically, as Figure 3 , Figure 4 , Figure 3 The SAG-Net network structure involved in the embodiment scheme is shown in the following figure, Figure 4For the structure diagram of the feature extraction module involved in the embodiment, the network realizes semantic segmentation of conductor, tower and background point cloud through point coding, downsampling, feature extraction and decoding process, wherein the processing flow of each module in the network is as follows: Optionally, in the embodiment, the step of obtaining the structure embedding feature through the structure prior embedding module comprises: Based on the three-dimensional coordinates, laser reflection intensity and neighborhood point cloud distribution of each point, the local point cloud density, local linear structure response and vertical structure response of each point are calculated; The three-dimensional coordinates, laser reflection intensity, local point cloud density, local linear structure response and vertical structure response are spliced to obtain enhanced input features; The enhanced input features are mapped to a preset feature channel dimension by a multilayer perception machine, and the structure embedding features are obtained.

[0058] The processing flow of the structure prior embedding module comprises: For any point in the input point cloud , first, construct its structure prior feature. The feature is used to describe the linear degree, vertical distribution feature and local density feature of the point in the local neighborhood. The enhanced input feature of point is expressed as: ;

[0059] Wherein, represents the coordinate value of the i-th point along the horizontal axis (cable span direction) in the local line coordinate system (or the original space coordinate system), represents the coordinate value of the i-th point along the vertical axis (transverse direction perpendicular to the line direction) in the local line coordinate system, represents the coordinate value of the i-th point along the vertical axis (vertical direction, i.e. elevation direction) in the local line coordinate system; represents the laser reflection intensity, represents the local point cloud density, represents the local linear structure response, represents the vertical structure response.

[0060] The local linear structure response is used to enhance the elongated continuous feature of the conductor point cloud. Let point be the neighborhood covariance matrix of point ;

[0061] The local linear structure response is defined as: ;

[0062] wherein, To prevent the denominator from being zero. The wire point cloud usually extends along a main direction, so it has a higher value; the vegetation and ground background points are more scattered, and the value is usually lower.

[0063] Vertical structure response For auxiliary identification of tower point cloud. The tower point cloud usually has obvious vertical extension and local lattice characteristics, so the angle relationship between the main direction of the point cloud neighborhood and the vertical direction can be constructed: ;

[0064] wherein, represents the unit vector of the main direction of the point neighborhood, represents the unit vector of the vertical direction. When the local structure is closer to the vertical distribution, is larger.

[0065] Subsequently, the enhanced input features are dimensioned by a multi-layer perception to obtain initial structure embedding features: ;

[0066] wherein: ;

[0067] represents the number of feature channels after dimensioning.

[0068] Further, in the embodiment, the processing steps of the encoding module include: Through the local geometry encoding branch, the farthest point sampling is performed on the point cloud corresponding to the structure embedding features of the current layer to obtain a low-resolution point set; Each sampling point in the low-resolution point set is established into a preset number of neighbor sets, and the local geometry features are output in combination with a local geometry feature mapping function and a neighborhood maximum pooling operation; Through the direction continuous encoding branch, the local span direction is determined according to the tower space position or the line direction, and the projection coordinates of each sampling point along the local span direction are calculated; The laser point cloud data is divided into direction strips according to the projection coordinates, the point features in each direction strip are aggregated by pooling to obtain strip-level features, and the one-dimensional continuous feature propagation is performed on the strip-level features of adjacent strips along the strip arrangement direction. The propagated strip-level features are fed back to the points in the corresponding strip to obtain the direction continuous enhanced features; The local geometry features and the direction continuous enhanced features are spliced along the feature channel dimension through the structure gate fusion unit to obtain spliced features; The splicing feature is input into a gated multi-layer perception and mapped to a preset interval combined with an activation function to determine a gating weight; The local geometric feature and the direction continuous enhancement feature are weighted by the gating weight, and the two weighted results are added element by element to obtain the fused point feature of each layer.

[0069] (1) The processing flow of the local geometric coding branch includes: The local geometric coding branch is used to extract the local spatial structure feature in the neighborhood of the point cloud. For the i-th layer, the input feature is: ;

[0070] wherein, represents the number of point cloud points of the i-th layer, represents the feature channel number of the i-th layer. First, the point cloud is sampled to obtain a point set with lower resolution and its feature:

[0071] ; ;

[0072] Then, a neighborhood set is established for each sampling point, and a relative position code is constructed: ;

[0073] wherein, represents the reference point (anchor point / sampling point) being processed, represents the neighborhood point of the reference point, represents the neighborhood set established for the reference point. The local geometric coding feature can be represented as: ;

[0074] wherein, represents a neighborhood max-pooling operation, represents a local geometric feature mapping function, represents the point feature vector of the neighborhood point in the i-th layer, represents the Euclidean distance of the relative position code, which is a scalar reflecting the spatial distance from the neighborhood point to the reference point, represents the local linear structure response of the neighborhood point, represents the vertical structure response of the neighborhood point. This branch focuses on learning the local morphological differences of the point cloud, so that the network can distinguish the slender structure of the conductor, the vertical structure of the tower, and the scattered structure of the vegetation / background. ​​​

[0075] The processing flow of the direction continuous encoding branch includes: Since the conductor extends continuously along the span direction in the transmission line, the conductor point cloud still has obvious direction continuity even if there is local point cloud missing. Therefore, the SAG-Net sets a direction continuous encoding branch for enhancing the continuous expression of the conductor points in the long distance direction.

[0076] First, the local direction unit vector is estimated according to the spatial position of the tower or the overall trend of the line . For any point , the projection coordinates of the point along the direction are calculated: ;

[0077] The point cloud is divided into several direction strips according to ;

[0078] wherein represents the th direction strip, and represents the th projection coordinate. The point features in each strip are aggregated to obtain the strip-level continuous features: ;

[0079] wherein may be any one of average pooling, maximum pooling or weighted pooling, represents the point feature vector of the reference point currently being processed in the th layer.

[0080] Subsequently, one-dimensional continuous feature propagation is performed between adjacent strips: ;

[0081] represents a sequence feature mapping multi-layer perception, and the strip-level continuous features are returned to the corresponding points to obtain the direction continuous enhanced features: ;

[0082] The branch is used to strengthen the continuity of the conductor along the direction and reduce the influence of local occlusion, point cloud sparsity and vegetation interference on the conductor segmentation result. Unlike ordinary point cloud segmentation networks, the branch utilizes the trend prior of the transmission line, so that the network is more suitable for serving the subsequent conductor sag fitting.

[0083] The processing flow of the structure gate fusion unit includes: ​The local geometry feature and the direction continuous enhancement feature correspond to a gating weight respectively, and after the gating weight of the local geometry feature is calculated, the gating weight of the corresponding direction continuous enhancement feature is 1 minus the gating weight of the local geometry feature.

[0084] The local geometry feature obtained by the local geometry encoding branch The feature direction continuous enhancement feature obtained by the direction continuous encoding branch pays more attention to the difference of local morphology. The feature direction continuous enhancement feature obtained by the direction continuous encoding branch pays more attention to the continuity of the conductor along the span direction. In order to adaptively fuse the two types of features, a structure gating fusion unit is introduced.

[0085] Firstly, the two types of features are spliced: ;

[0086] Then the gating weight is calculated: ;

[0087] wherein, Sigmoid activation function is denoted as, The local geometry feature mapping function is denoted as.

[0088] The fused point feature is: ;

[0089] wherein, Element-wise multiplication is denoted as.

[0090] When the point is more consistent with the slender continuous structure of the conductor, the network can increase the weight of the direction continuous feature; when the point is closer to the tower or the complex background structure, the network can increase the weight of the local geometry feature. In this way, SAG-Net can form a more stable structure distinction between the conductor, the tower and the background.

[0091] Further, in the embodiment, the step of obtaining the decoded point-level feature through the point-level decoding module comprises: Obtaining the input feature of the current decoding layer, and processing the input feature of the current decoding layer by nearest neighbor interpolation to obtain an interpolated feature, wherein the number of feature points of the interpolated feature is restored to be consistent with the number of fused point features corresponding to the current decoding layer; Splicing the interpolated feature and the fused point feature output by the corresponding encoding stage along the feature channel dimension, and performing feature fusion through a decoding multi-layer perception to obtain the point-level feature.

[0092] Specifically, after the multi-layer structure feature encoding is completed, SAG-Net restores the low-resolution semantic feature to the original point cloud resolution through the point-level decoding module. For the first decoding input feature of the i-th decoding layer, the interpolation is performed on the input feature of the i-th decoding layer to obtain an interpolated feature, and the number of feature points of the interpolated feature is restored to be consistent with the number of fused features corresponding to the i-th decoding layer. layer decoding input feature Firstly, the point number is recovered by nearest neighbor interpolation or inverse distance weighted interpolation: ;

[0093] Then the point-level features are obtained by fusing the shallow features reserved in the corresponding encoding stage: ;

[0094] wherein, denotes the point feature interpolation operation, denotes the structure-enhanced feature output by the corresponding encoding layer, denotes the decoding multi-layer perception, denotes the interpolated feature obtained after performing the interpolation operation on the input feature of the current decoding layer.

[0095] Finally, the decoded point-level features are input into the semantic prediction head to obtain the probability that each point belongs to the conductor, tower or background: ;

[0096] wherein, denotes the Softmax activation function, which is used to normalize the original classification scores output by the classification multi-layer perception into a probability distribution, denotes the classification multi-layer perception; ;

[0097] The point-level semantic result is obtained according to the maximum probability category: ;

[0098] wherein, , denotes the category corresponding to the maximum probability, denotes the semantic prediction probability, denotes the conductor category, which is the metal conductor carrying current in the transmission line, and its geometric feature is an elongated continuous structure, which is represented as a linear distribution extending along the span direction in the point cloud; denotes the tower category, which represents the transmission tower supporting the conductor, and its geometric feature is a vertical extension of the lattice structure, which is represented as a combination of pole pieces distributed along the vertical direction in the point cloud; denotes the background category, which represents the vegetation, ground, buildings and other point cloud data in the line corridor that are not conductors or towers, and its geometric feature is scattered distribution without obvious linear or vertical structure regularity.

[0099] Finally, the conductor point cloud and the tower point cloud are extracted based on the point-level semantic result, and the conductor point set and the tower point set are extracted according to the semantic prediction result output by the SAG-Net: ; ;

[0100] in, This represents the set of candidate point clouds for the conductor. This represents the set of candidate points for poles.

[0101] In addition, SAG-Net needs to be trained. During the training phase, SAG-Net uses point-level cross-entropy loss to supervise the semantic labels of three classes: conductors, towers, and background. ;

[0102] in, For the first The true category label of each point, For the network to predict which one belongs to the first The probability of a class This indicates the number of points to input.

[0103] To enhance the continuity of the traverse point cloud, a traverse continuity constraint can be further introduced. For traverse points that are spatially adjacent and located within the same or adjacent stripes, their traverse class probabilities should remain consistent. This constraint can be expressed as: ;

[0104] in, Represents the set of adjacent pairs of points. Point The probability of belonging to the wire category. express The probability of belonging to the wire category. This represents the adjacency weight between point pairs. If two points are spatially close and continuous along a direction, then... Larger.

[0105] The final training loss is: ;

[0106] in, These are the weights for continuity constraints.

[0107] This training objective does not require centerline offset annotations or point-level confidence supervision labels. SAG-Net is only responsible for outputting point-level semantic results for conductors, towers, and the background; subsequent point-level fitting confidence is calculated jointly by conductor class probabilities, local linear structures, point cloud density, and catenary fitting residuals.

[0108] Step S30: determining span ranges between adjacent towers according to the tower point cloud and the tower spatial positions, extracting conductor candidate point clouds from the conductor point cloud in each span range, performing single-conductor separation on the conductor candidate point clouds to obtain single-conductor point sets.

[0109] In the embodiment, the prior tower spatial position refers to determining the accurate position coordinates of each base tower in the three-dimensional space according to the spatial distribution of the tower point cloud. The tower sorting is to arrange all the identified towers in the order from front to back along the direction of the transmission line to establish a tower sequence. The span refers to the horizontal distance between two adjacent base towers and the spatial region within the distance range, which is the basic spatial unit of the transmission line. The conductor candidate point cloud refers to the set of points belonging to the conductor category extracted from each span spatial range. These points may contain point cloud data of multiple conductors.

[0110] Specifically, the base towers are positioned and sorted according to the obtained spatial positions of the tower point cloud. First, the spatial clustering based on density is performed on the tower point cloud, and the spatially aggregated tower points are divided into independent tower point clusters, and the geometric center of each tower point cluster is calculated as the spatial position coordinates of the base tower. Then, the projection coordinates of each base tower are sorted according to the direction of the line, and the tower sequence from front to back along the line is obtained. Next, the space range between two adjacent base towers in the tower sequence is determined as a span, and all conductor points falling within the span spatial range are extracted from the conductor point cloud as the conductor candidate point cloud of the span.

[0111] Optionally, in the embodiment, the step of extracting the conductor candidate point cloud from the conductor point cloud in each span range, performing single-conductor separation on the conductor candidate point cloud to obtain a single-conductor point set, includes: In the conductor candidate point cloud in each span, a spatial adjacency relationship is established according to the comparison result of the three-dimensional spatial distance between each point and the preset neighborhood radius threshold; Taking the spatial adjacency relationship as the basis for connection, performing connected clustering on all conductor candidate points to determine an initial clustering cluster; Calculating the principal direction of the initial clustering cluster through principal component analysis, and dividing the point cloud into at least two slices along the principal direction, and counting the distribution peaks of the point cloud in the horizontal and vertical directions in each slice; Segmenting the point cloud at the corresponding slice position into a preset number of sub-clusters along the gap between the distribution peaks, and iteratively performing the segmentation operation until the horizontal and vertical distributions of each sub-cluster in the principal direction are single peaks, and finally determining the sub-cluster as a single-conductor point set corresponding to a spatially independent continuous conductor.

[0112] On this basis, geometric clustering is performed on the conductor candidate point cloud in each span to separate single conductors: in the conductor candidate point cloud in each span, a spatial adjacency relationship is established according to the comparison result of the three-dimensional spatial distance between points and the preset neighborhood radius threshold value--if the three-dimensional spatial distance between two points is less than the preset neighborhood radius threshold value, it is determined that the two points are spatially adjacent points; taking the spatial adjacency relationship as the basis for connectivity, all conductor candidate points are subjected to connected clustering, and points directly or indirectly connected through the spatial adjacency relationship are classified into the same initial clustering cluster, and points without spatial adjacency relationship are classified into different initial clustering clusters; the principal direction of each initial clustering cluster is calculated through principal component analysis, and the point cloud is divided into multiple slices along the principal direction, and the distribution peaks of the point cloud in the horizontal and vertical directions of each slice are counted; for the slices with multiple distribution peaks, the point cloud at the corresponding slice position is segmented into multiple sub-clusters along the gap between the peaks, and the slice segmentation operation is iteratively performed until the horizontal and vertical distributions of each sub-cluster in the principal direction are single peaks, and each final sub-cluster is determined as a single conductor point set corresponding to a spatially independent continuous conductor.

[0113] Optionally, the spans are automatically divided based on the tower point cloud identification result, such as Figure 5 , Figure 5 For the implementation example scheme, the unmanned aerial vehicle laser radar collection and span division schematic diagram is shown in FIG. 1. The unmanned aerial vehicle collects point clouds along the direction of the power transmission line, and divides the spans based on the positions of adjacent towers.

[0114] According to the tower point cloud and the known tower spatial position, the towers are positioned and sorted. Let the identified tower set be: ;

[0115] wherein, represents the first base tower. After sorting the towers according to the line direction, the spatial range between the two adjacent base towers and defines a span: ;

[0116] The corresponding span point cloud is: ;

[0117] The span length is calculated according to the spatial positions of adjacent towers, or is determined with the aid of known tower coordinates.

[0118] Optionally, based on geometric clustering to separate single conductor point clouds, for each span point cloud , the points belonging to the conductor category are extracted to obtain the conductor candidate point cloud in the span: ;

[0119] This embodiment focuses on icing measurement for a single conductor. For multiple conductors spatially separated within the same span, a clustering method based on spatial continuity is used to separate them, resulting in several sets of individual conductor points: ;

[0120] in, Indicates the first Within the first gear range A single wire.

[0121] This application employs Euclidean clustering, spatial continuity analysis, and principal direction consistency constraints to separate individual conductors from the conductor point cloud within the span. This approach can also be replaced by instance segmentation networks, graph clustering methods, or conductor tracing methods based on tower suspension point topology.

[0122] Among these methods, instance segmentation networks require additional instance-level annotations, resulting in high data construction costs; graph clustering methods are sensitive to point cloud density and adjacency thresholds; and methods based on hanging point topology rely on the quality of hanging point identification. This application preferably adopts a geometric clustering approach, which does not require distinguishing between the three-phase electrical attributes (A, B, and C) or additional instance annotations, making it more suitable for the single conductor icing measurement scenario of this application.

[0123] Automatic span division was achieved through the automatic positioning and sorting of tower point clouds, decomposing long-distance transmission lines into several independent span units without manual intervention. Precise separation of multiple conductors within the same span was achieved through connected clustering based on spatial adjacency and slicing based on the main direction. Connected clustering utilizes the physical property of conductors being spatially separated to initially separate different conductors, while slicing based on the main direction further solves the problem of connected clustering potentially misclassifying multiple conductors into the same cluster when they are spatially close. These two steps of clustering and separation ensure that each single conductor point set corresponds to one and only one physical conductor, providing accurate input data for subsequent centerline fitting and sag calculation for each conductor.

[0124] Step S40: Calculate the point-level fitting confidence of each traverse point based on the traverse category probability, local linearity, local point cloud density, and fitting residual confidence of each traverse point in the single traverse point set, and perform weighted catenary fitting on the single traverse point set based on the point-level fitting confidence to obtain the three-dimensional centerline of the corresponding traverse.

[0125] In this embodiment, the three-dimensional centerline refers to the center trajectory curve that describes the geometric direction of the conductor in three-dimensional space. This embodiment uses a catenary model to describe the vertical coordinates of the conductor along the span direction and a polynomial model to describe the coordinates of the conductor along the transverse direction.

[0126] Specifically, for each single-wire point set, firstly, the point-level fitting confidence of each point is calculated. For each point in the single-wire point set, the wire category probability output by the sagittal perception point cloud segmentation network in step S20, the local linearity calculated based on the eigenvalue of the neighborhood covariance matrix, the local point cloud density calculated based on the number of neighborhood points, and the distance residual of the point to the initial fitting curve are obtained; the above four factors are weighted and summed according to the preset weight coefficients to obtain the point-level fitting confidence of the point. Among them, the fitting residual confidence is calculated according to the distance of the point to the initial fitting curve by a Gaussian kernel function, and the smaller the distance, the higher the confidence.

[0127] Then, the random sample consensus algorithm is used to iteratively fit the single-wire point set: a minimum sample subset is randomly extracted from the single-wire point set, an initial curve model is fitted based on the extracted subset, the distance of all points to the initial curve is calculated, the points with a distance less than a preset threshold are marked as inliers and are included in the inlier set, and the size of the current inlier set and the fitting residual are recorded; the above random sampling and fitting process is repeated a preset number of times, and the inlier set corresponding to the model with the most inliers and the smallest fitting residual is selected as the final inlier set, thereby eliminating outliers and misclassified points.

[0128] Finally, in the local line coordinate system, the catenary model is used to describe the vertical coordinates of the wire along the span direction, and the polynomial model is used to describe the coordinates of the wire along the transverse direction, and the point-level fitting confidence calculated in step S40 is used as the weight to perform weighted least squares fitting on the inlier set: a weighted residual sum of squares objective function is constructed with the catenary parameters and the polynomial parameters as unknowns, the derivative of the objective function is taken and set to zero, and the model parameters that minimize the weighted residual sum of squares are solved, thereby reconstructing the three-dimensional center line of the wire. Among them, the point-level fitting confidence is a quantitative evaluation index of the reliability of each point in the single-wire point set, which is used to control the weight of different points in subsequent curve fitting. The confidence reflects the credibility of a point as a fitting sample of the wire center line: the higher the wire category probability, the more obvious the local linear structure, the more reasonable the neighborhood point cloud density, and the smaller the distance residual of the point to the preliminary fitting curve, the higher the confidence. The inlier set refers to the subset of points filtered from the original single-wire point set by the random sample consensus algorithm, which has consistency with the catenary model and eliminates outliers and misclassified points. The random sample consensus algorithm is an iterative method for estimating mathematical model parameters by repeatedly random sampling and model verification, which can effectively handle the case where a large number of outliers are included in the data.

[0129] Through the comprehensive calculation of the point-level fitting confidence, high-quality wire points (high classification probability, obvious linear structure, reasonable density, and small residual) obtain high weight in fitting, and low-quality points (misclassified points, outliers, and edge sparse points) obtain low weight, thereby effectively suppressing the negative impact of noise points on the center line fitting; the random sample consensus algorithm is iteratively screened to further remove outliers and misclassified points, thereby ensuring that the point set participating in the fitting has high consistency; the catenary model is used to fit the vertical coordinates of the wire, so that the reconstructed center line conforms to the natural suspension form of the wire under the action of gravity and tension, thereby having higher physical rationality and geometric stability than ordinary polynomial fitting or spline fitting, and laying a geometric foundation for the accurate calculation of the sag.

[0130] Optionally, in the embodiment, according to the wire category probability, local linearity, local point cloud density, and fitting residual confidence of each wire point in the single-wire point set, the point-level fitting confidence of each wire point is calculated, and the single-wire point set is weighted catenary fitting based on the point-level fitting confidence, to obtain the three-dimensional center line of the corresponding wire. The wire category probability, local linearity, and local point cloud density of each wire point in the single-wire point set are obtained. The fitting residual confidence is calculated according to the distance of the wire point to the initial fitting curve and the scale parameter. The point-level fitting confidence of each wire point is weighted calculated according to the wire category probability, local linearity, local point cloud density, and fitting residual. The random sample consensus algorithm is used to randomly sample and fit the initial curve for the single-wire point set, to obtain an inner point set. In the local line coordinate system, the catenary model is used to describe the vertical coordinates of the wire along the span direction, and the lateral model is used to describe the coordinates of the wire along the lateral direction. The inner point set is weighted least squares fitted with the point-level fitting confidence as the weight, and the catenary model parameters and the lateral model parameters are solved, to obtain the three-dimensional center line of the wire.

[0131] For each single-wire point set , the point-level fitting confidence is constructed, which is used to assign different weights to different points in the subsequent catenary fitting. The point-level fitting confidence is calculated by means of the semantic classification probability, local linearity, local point cloud density, and fitting residual.

[0132] For the wire point , the confidence is defined as: ;

[0133] wherein, is the wire category probability output by the SAG-Net, is the normalized local linearity, is the normalized local point cloud density, is the fitting residual confidence, is the weight coefficient.

[0134] where the fitting residual confidence can be calculated according to the preliminary fitting curve: ;

[0135] where, denotes the distance of the point to the preliminary fitting curve, is the scale parameter.

[0136] When the derivative classification probability of a certain point is high, the local linear structure is obvious, the neighborhood density is reasonable, and the residual distance to the preliminary curve is small, the confidence is high; when the certain point is a misclassified point of vegetation, an outlier point, a shelter edge point, or a sparse abnormal point, the confidence is low.

[0137] Based on the RANSAC (Random Sample Consensus) and catenary model fitting of the three-dimensional center line of the conductor, for each single conductor point set , first, the RANSAC method is used to remove outliers and misclassified points to obtain an inner point set: ;

[0138] RANSAC repeatedly samples the conductor points and fits the initial curve model, judges the inner point set according to the distance of the points to the curve, and finally selects the model with more inner points and smaller fitting residual as the initial model.

[0139] After obtaining the inner point set, the catenary model is used to fit the three-dimensional center line of the conductor. Let denote the coordinate along the span direction, denote the transverse coordinate, denote the vertical coordinate. The conductor center line is represented as: ;

[0140] where the vertical curve is fitted using the catenary model: ;

[0141] where, is the catenary fitting parameter.

[0142] The transverse curve can be fitted using a low-order polynomial or a spline curve: ;

[0143] wherein, is a transverse curve fitting coefficient, is a polynomial order.

[0144] The catenary fitting adopts a weighted least square method, and the objective function is: ;

[0145] wherein, denotes a center line model parameter set, denotes a point level fitting confidence, denotes a point distance to the fitted center line.

[0146] Through the above steps, the three-dimensional center line of the i-th conductor in the j-th span is obtained: ; ;

[0147] The purpose of this step is to generate a center reference line that is geometrically smooth, structurally continuous, and consistent with the conductor catenary shape, which is used for subsequent sag calculation. Compared with directly estimating the ice thickness by using the outer contour of the conductor point cloud, the equivalent ice thickness inversion by the center line and the sag change has stronger adaptability to local missing of the point cloud, incomplete outer contour of the conductor, and irregular ice layer shape.

[0148] Optionally, the application adopts a combination of RANSAC denoising and catenary model weighted fitting to reconstruct the three-dimensional center line of the conductor. This method can be replaced by three-dimensional polynomial fitting, B-spline curve fitting, Bezier curve fitting, or multi-segment spline fitting method. The above methods can also realize the modeling of the spatial trajectory of the conductor. However, the power transmission conductor usually presents a catenary shape under the action of gravity and tension, and the catenary model is more consistent with the actual force and spatial shape of the conductor. Compared with ordinary polynomials or spline curves, the catenary fitting adopted by the application can improve the physical rationality and stability of the sag calculation, and is especially suitable for long span or high difference of hanging points of power transmission line scenes.

[0149] Step S50: According to the spatial adjacency relationship between the two ends of the three-dimensional center line and the corresponding tower point cloud, the hanging points at the two ends of the conductor are identified, and the connecting line of the two hanging points is taken as a reference chord line. The maximum vertical difference of the three-dimensional center line relative to the reference chord line is calculated to obtain the current sag.

[0150] ​In the embodiment, the hanging point is a fixed connection point of the conductor on the tower, that is, the connection position of the conductor and the insulator string or the tower cross arm, which is represented as the intersection area of the conductor center line end and the tower point cloud in space. The reference chord is a straight line connecting between the two hanging points of the conductor, which is used as a reference line to measure the sag degree of the conductor in the sag measurement. The current sag refers to the maximum vertical sag distance of the conductor center line relative to the connecting line between the two hanging points under the current icing state, which is a key geometric parameter reflecting the load state of the conductor. According to the basic definition of the power transmission project, the sag is the vertical distance between the connecting line of the hanging point and the lowest point of the conductor, and the difference is positive when the conductor center line is below the connecting line of the hanging point.

[0151] Specifically, according to the three-dimensional center line reconstructed in step S40, within the preset search radius of the center line end, the hanging point candidate is determined in combination with the tower cross arm region, the conductor end direction, the endpoint projection distance and the nearest neighbor tower point set, and the one with the minimum distance residual between the fitted center line end and the candidate hanging point is taken as the hanging point. For any one conductor within a span, the point cloud data of the towers at both ends of the span is obtained, within the preset spatial search range of the starting end of the three-dimensional center line, the three-dimensional center line point adjacent to the tower point cloud in space within the range is searched, and the coordinate of the point is determined as the spatial coordinate of the first end hanging point. In the same way, the spatial coordinate of the second end hanging point is determined at the terminal end of the three-dimensional center line. After determining the spatial coordinates of the two end hanging points, the main body takes the connecting line of the two end hanging points as the reference chord: the coordinates and vertical coordinates of the first end hanging point in the span direction are obtained, the coordinates and vertical coordinates of the second end hanging point in the span direction are obtained, and the vertical coordinates of the reference chord at each span direction position are calculated according to the linear interpolation formula. At the same time, the vertical coordinates of the three-dimensional center line at the same span direction position are obtained. Finally, the difference between the vertical coordinates of the reference chord and the vertical coordinates of the three-dimensional center line is calculated, and the maximum value of the difference between the two end hanging points is taken as the current sag under the current state.

[0152] The hanging point is automatically identified through the spatial adjacency relationship between the three-dimensional center line end and the tower point cloud, without manual annotation or dependence on tower design parameters, so that the hanging point identification process is completely automated and can be repeatedly executed. The sag is determined by taking the connecting line of the two end hanging points as the reference chord and calculating the maximum vertical sag distance, which strictly follows the definition standard of the sag in the power transmission project, so that the sag calculation result has a clear physical meaning and engineering interpretability. By searching for the maximum difference in the entire span range instead of only searching for the lowest point, it is ensured that the maximum sag position can be accurately captured even if the conductor center line has local fluctuations.

[0153] Optionally, in this embodiment, the steps of determining candidate hanging points based on the three-dimensional centerline, within a preset search radius at the end of the centerline, and combining the tower crossarm area, conductor end direction, endpoint projection distance, and nearest neighbor tower point set, and selecting the hanging point with the smallest residual distance between the end of the fitted centerline and the candidate hanging point, and using the connecting line between the two hanging points as a reference chord, include: Determine the spatial coordinates of the two hanging points, and use the line connecting the two hanging points as the reference chord. Calculate the vertical coordinates of the reference chord at each span direction; Obtain the vertical coordinates of the three-dimensional centerline at the same span direction, calculate the difference between the vertical coordinates of the reference chord and the vertical coordinates of the three-dimensional centerline, and take the maximum value of the difference between the two hanging points as the current sag.

[0154] For the center line of each conductor The attachment point is automatically identified based on the spatial adjacency between its two ends and the point cloud of the tower. Let the first... The towers at both ends of each span are and Near the tower Within the range of the conductor's end, search for the spatial connection point between the conductor's centerline and the tower crossarm area or conductor connection area to obtain the first end attachment point: ;

[0155] Near the tower Within the range of the conductor end, search for the second end attachment point: ;

[0156] Determine the hanging points at both ends and Then, the line connecting the two suspension points is used as the reference chord. For the center line of the conductor... any position on The vertical coordinate of the line connecting the hanging points at this location is: ;

[0157] in, and These represent the hanging points. and Coordinates in the distance direction.

[0158] The vertical coordinates of the three-dimensional centerline at the same location are: ;

[0159] In the current state, the conductor sag is defined as the maximum vertical downward distance of the conductor's centerline relative to the line connecting the suspension point; this is the current sag. ;

[0160] Step S60: Determine the sag increment based on the current sag and the reference sag; calculate the equivalent horizontal tension based on the reference sag, span length, and conductor foundation information under historical ice-free conditions; and invert the additional load per unit length caused by icing based on the sag increment, span length, and equivalent horizontal tension.

[0161] In this embodiment, the reference sag refers to the sag value of the same transmission line, span, and conductor under historical ice-free conditions, reflecting the normal spatial morphology of the conductor under no additional ice load. The sag increment refers to the difference between the sag under current icing conditions and the reference sag under historical ice-free conditions, reflecting the degree of change in the conductor's spatial morphology caused by icing load. The equivalent horizontal tension refers to equating the complex stress state of the conductor under its own weight and tension to a horizontal tension value along the span direction, used to establish a quantitative relationship between sag and load. According to catenary theory, given the load per unit length of the conductor, span length, and sag, the equivalent horizontal tension can be calculated back through mechanical relationships.

[0162] Further, in this embodiment, step S60 includes: The sag increment is determined based on the current sag and the reference sag, and the equivalent horizontal tension is calculated based on the conductor foundation information, span length and the reference sag under historical ice-free conditions. The icing mass per unit length is determined based on the sag increment, span length, and equivalent horizontal tension. The icing thickness is obtained by calculating the icing mass per unit length, the bare conductor radius, and the icing density using an equivalent circular model.

[0163] Specifically, such as Figure 6 As shown, Figure 6 This is a schematic diagram of conductor centerline reconstruction and sag calculation involved in this embodiment. Historical ice-free point cloud data of the same transmission line, the same span, and the same conductor under historical ice-free conditions are obtained. The same processing flow described in steps S10 to S50 is sequentially executed on the historical ice-free point cloud data, including point cloud preprocessing, semantic segmentation, span division, single conductor clustering, three-dimensional centerline fitting, hanging point identification, and sag calculation to obtain the reference sag of the corresponding conductor.

[0164] Then, the difference between the current sag and the reference sag is calculated to obtain the sag increment. Based on this, the equivalent horizontal tension is calculated by back-calculating the conductor unit length load, span length, and reference sag under historical ice-free conditions, according to the mechanical relationship between catenary sag and horizontal tension. The conductor unit length load is obtained by multiplying the unit length mass corresponding to the conductor type by the acceleration due to gravity, and the span length is determined by the coordinate difference between the two suspension points in the span direction.

[0165] By using the same line history ice-free state point cloud to calculate the reference sag, instead of relying entirely on the theoretical design sag, the reference sag can truly reflect the actual operation form of the line in the ice-free state, thereby reducing the deviation between pure theoretical calculation and actual state; by comparing the current sag with the reference sag, the systematic error introduced by the difference in data sources and processing methods is eliminated; by calculating the equivalent horizontal tension from the reference sag, the subsequent inversion of ice-coating load is based on the actual stress state of the line, thereby improving the accuracy of ice thickness calculation.

[0166] Optionally, the reference sag is calculated based on historical ice-free point cloud: unmanned aerial vehicle laser point cloud data collected in the historical ice-free state of the same power transmission line and the same span is obtained. The same processing flow is performed on the historical ice-free point cloud, including point cloud preprocessing, tower identification, span division, conductor extraction, single conductor separation, catenary fitting, hanging point identification, and sag calculation, to obtain the reference sag of the corresponding conductor , which represents the sag of the th span and the th conductor in the historical ice-free state. The difference between the current state sag and the reference sag is: ;

[0167] wherein, represents the conductor sag increment.

[0168] Optionally, the reference sag is calculated based on historical ice-free point cloud of the same line. This method can also be replaced by theoretical ice-free sag calculated from line design parameters, historical sag recorded by fixed online monitoring devices, or reference sag obtained by manual measurement. Among them, the design parameter calculation method is convenient to implement, but is easily affected by construction errors, long-term operation state changes and tension changes of the line; the data of the fixed monitoring device is reliable, but the coverage is limited; manual measurement is low in efficiency. The historical ice-free point cloud is used as reference data in the present application, so that the current sag and the reference sag are both derived from the same unmanned aerial vehicle point cloud data and the same processing flow, thereby reducing the error caused by the difference in data sources.

[0169] Optionally, temperature fitting correction can also be performed. In the case of having on-site temperature data, optional temperature correction can be performed on the sag increment. Let the temperature collected in the historical ice-free state be , and the temperature collected in the current ice-coating state be , then the sag change caused by temperature can be represented as: ;

[0170] wherein, represents the temperature sag correction coefficient, which can be fitted from the multi-period historical ice-free point cloud data, on-site temperature data and sag change data of the same line.

[0171] The temperature-corrected sag increment is: ;

[0172] If reliable temperature fitting coefficients are not available, or temperature correction is not enabled, then: ;

[0173] It is noted that temperature correction is an optional step of the present application, and is not a necessary condition for implementing the present application.

[0174] Determine the equivalent horizontal tension and inverse the icing-induced load: Determine the equivalent horizontal tension according to the line account parameters and the historical non-icing state sag calculation results. The line account parameters include one or more of the following: conductor type, conductor radius, unit length mass, design tension, elastic modulus, cross-sectional area, and span length.

[0175] Inverse calculation is performed according to the historical non-icing state. Let the unit length load of the conductor in the non-icing state be , the span length be , and the historical non-icing reference sag be , then the equivalent horizontal tension can be approximately expressed as: ;

[0176] The unit length load induced by icing is obtained by inverse calculation from the corrected sag increment. In a general form, it is expressed as: ;

[0177] wherein is the unit length load induced by icing, is the set of conductor account parameters, is the load inverse function established from the mechanical relationship of the conductor.

[0178] In a simplified implementation, when the sag change satisfies the engineering approximation condition, the following can be used: ;

[0179] This formula is used to convert the sag increment to the unit length load of the conductor, represents the square of the span length, represents the temperature-corrected sag increment.

[0180] The present application inverses icing load based on sag increment, span length, conductor account parameters and equivalent horizontal tension. The process can use a simplified parabolic relationship, or a catenary state equation or a conductor mechanics state equation for inversion. In addition, tension sensors, finite element simulation or historical data regression models can also be used to estimate icing load. However, tension sensors require the installation of fixed monitoring devices, finite element methods are complex to calculate, and data-driven models are strongly dependent on the number of samples. The present application preferably uses an inversion method combining sag increment and conductor mechanics relationship, which does not require the installation of additional sensors on the conductor and is more suitable for unmanned aerial vehicle mobile ice observation scenarios.

[0181] Alternatively, the unit length icing mass is determined according to the sag increment, span length and equivalent horizontal tension, and the unit length icing mass, conductor bare radius and icing density are calculated based on the circular ring equivalent model to obtain the icing thickness, which specifically includes: In the present embodiment, the unit length icing mass refers to the mass of icing per unit length of conductor, which is a quantitative representation of icing load. The circular ring equivalent model is an engineering approximation method that simplifies the actual possibly uneven icing into a uniform circular ring cross section concentric with the conductor. The core idea is to use a uniform circular ring with the same cross section to equivalently represent the mass and cross-sectional area of the actual icing, thereby simplifying the complex icing shape problem into a geometric problem that is easy to calculate. In this model, the conductor bare radius is the inner diameter of the circular ring, the sum of the conductor radius and the icing thickness is the outer diameter of the circular ring, and the unit length icing mass is equal to the icing density multiplied by the circular ring cross-sectional area. The icing thickness refers to the radial distance from the surface of the conductor to the outer surface of the ice layer under the equivalent circular ring model, and is the core index for evaluating the severity of icing in engineering.

[0182] By inversing the unit length icing load through the sag increment, and then calculating the equivalent icing thickness through the circular ring equivalent model, the indirect measurement of icing thickness is achieved. Compared with the traditional method of directly measuring the outer contour of the ice layer, the present application does not require the outer surface of the icing conductor to be completely scanned, nor does it require the ice layer to be uniform and regular. Even if the ice layer distribution is uneven, the point cloud is sparse or the scanning angle is limited, as long as the conductor center line can be reconstructed through the point cloud and reliable sag increment can be obtained, the equivalent icing thickness can be inversely calculated. The circular ring equivalent model simplifies the actual possibly irregular icing shape into a uniform circular ring cross section concentric with the conductor, making the calculation of icing thickness have clear geometric meaning and engineering operability. The entire inversion process only depends on the geometric shape change of the conductor center line and the conductor account parameters, without the need for any sensors installed on the conductor, fully adapting to the non-contact measurement requirements in the unmanned aerial vehicle mobile inspection scenario.

[0183] Step S70: converting the unit length icing load into unit length icing mass, and calculating the equivalent icing thickness of the power transmission line to be measured based on the conductor bare radius, icing density and circular ring equivalent model.

[0184] Optionally, in the embodiment, the step of determining the unit length icing mass according to the sag increment, the span length and the equivalent horizontal tension, and calculating the unit length icing mass, the conductor bare radius and the icing density by the circular ring equivalent model to obtain the icing thickness, comprises: According to the sag increment, the span length and the equivalent horizontal tension, the unit length additional load caused by icing is calculated, and the formula is: ;

[0185] wherein, is the unit length additional load caused by icing, is the conductor account parameter set, is the load inversion function established by the conductor mechanics relationship; The unit length icing mass is obtained by dividing the unit length additional load by the gravitational acceleration; The conductor bare radius and the icing density are obtained, and the icing thickness is calculated based on the circular ring equivalent model, and the formula is: ;

[0186] wherein, represents the icing thickness of the m-th conductor in the n-th span, represents the conductor bare radius, which is obtained from the conductor type library or the line account, represents the icing density determined according to the ice form, represents the unit length additional load caused by icing. Specifically, the equivalent icing thickness is calculated based on the circular ring equivalent model to convert the unit length additional load into the unit length icing mass:

[0187] ; ;

[0188] wherein, is the gravitational acceleration.

[0189] The icing density is determined according to the ice form . The ice form includes one of glaze, rime, wet snow icing or mixed icing. Different ice forms correspond to different icing density parameters. The icing is equivalent to a circular ring ice layer uniformly coated outside the conductor. Assuming that the conductor bare radius is , and the equivalent icing thickness is , the unit length icing mass satisfies: ;

[0190] Therefore, the equivalent icing thickness is: ;​

[0191] Substituting, we get: ;

[0192] wherein, represents the equivalent ice thickness of the m-th conductor in the n-th span; represents the bare radius of the conductor, which can be obtained from the conductor type library or the line account book; represents the ice density determined according to the ice shape; represents the unit length additional load caused by ice. Optionally, the application adopts a circular ring equivalent model to equivalently convert the ice coating load into a circular ring ice layer uniformly coated on the outer periphery of the conductor, and to calculate the equivalent ice thickness. This method can also be replaced by an elliptical cross-section equivalent model, a non-uniform ice coating cross-section model or an empirical ice thickness conversion model. The elliptical cross-section or non-uniform cross-section model can more specifically describe the eccentric ice coating and the asymmetric ice shape, but requires more complete ice layer appearance or cross-section information. The application adopts the circular ring equivalent model, which can directly convert the ice coating load into the equivalent ice thickness that is easy to understand and apply in engineering, and is more suitable for span-level ice risk assessment.

[0193] In summary, the alternative solutions can all achieve the calculation of the equivalent ice thickness of the conductor of the power transmission line to a certain extent, but the application preferably adopts the combined scheme of "arc sag perception point cloud segmentation, geometric clustering single conductor separation, RANSAC and catenary fitting, historical ice-free arc sag comparison, ice coating load inversion and circular ring equivalent ice thickness calculation". This combined scheme can take into account the stability of the conductor point cloud extraction, the physical reasonableness of the conductor center line modeling and the engineering implementability in the unmanned aerial vehicle moving ice observation scene.

[0194] In the technical scheme provided in the embodiment, the conductor point cloud and the tower point cloud are stably extracted from the complex line corridor point cloud by the arc sag perception point cloud segmentation network, the single conductor point set is separated by span division and geometric clustering, the three-dimensional center line of the conductor is reconstructed by weighted fitting of the catenary model, the current arc sag is calculated by hanging point identification and arc sag definition, the arc sag increment is obtained by comparison with the historical ice-free point cloud, and the equivalent horizontal tension is calculated, and finally the additional load is inverted by the arc sag increment, and the equivalent ice thickness is calculated based on the circular ring equivalent model. It does not rely on the direct measurement of the ice contour, but indirectly inverts the influence of the ice on the mechanical state and spatial form of the conductor, effectively solving the engineering problem that the ice coating contour of the conductor is difficult to stably obtain under the condition of the unmanned aerial vehicle laser point cloud, providing a non-contact, span-level and engineering implementable ice thickness calculation method for mobile ice observation, ice risk assessment and ice patrol auxiliary decision of the power transmission line.

[0195] In the technical scheme provided in the embodiment, the conductor point cloud and the tower point cloud is stably extracted from the complex line corridor point cloud by the arc sag perception point cloud segmentation network; the single conductor point set is separated by span division and geometric clustering; the three-dimensional center line of the conductor is reconstructed by weighted fitting of the catenary model; the current arc sag is calculated by hanging point identification and arc sag definition; the arc sag increment is obtained by comparison with the historical ice-free point cloud; and the equivalent horizontal tension is calculated. Finally, the additional load is inverted by the arc sag increment, and the equivalent ice thickness is calculated based ​​

[0196] In addition, the ice thickness of the span stage is summarized and the grade output is output. For the target conductor in the same span, the equivalent ice thickness of a single conductor can be summarized as the equivalent ice thickness of the span stage: ;

[0197] Among them, The summary function can take the average value, maximum value or designated representative value of the target conductor.

[0198] In an embodiment of the application, the maximum value is used as the representative value of the risk of the span stage to ensure that the ice risk assessment is biased to the safe side: According to the equivalent ice thickness of the span stage The ice grade is output. Assuming that the ice grade threshold is and Then: When , it is judged as slight icing or no obvious icing; when , it is judged as moderate icing; when , it is judged as severe icing.

[0199] At the same time, a three-dimensional visualization result is output, and the ice thickness, ice grade and spatial position of different spans are marked in the point cloud scene.

[0200] Optionally, in the embodiment, after the step of converting the unit length additional load into unit length ice mass and calculating the equivalent ice thickness of the power transmission line to be measured based on the conductor bare radius, ice density and circular ring equivalent model, the method further comprises: According to the conductor point cloud integrity, the conductor segmentation quality, the single conductor clustering continuity, the three-dimensional center line fitting residual error, the hanging point recognition quality and the historical ice-free benchmark matching quality, the result confidence of the ice thickness calculation result is calculated; When the result confidence is lower than a preset threshold, a low confidence label is output; When the conductor hanging point is missing, the three-dimensional center line fitting fails or the historical ice-free benchmark cannot be matched, an uncalculated label is output.

[0201] In order to avoid misjudgment caused by missing conductor point cloud, failure of hanging point recognition, too large catenary fitting residual error or unreliable historical ice-free benchmark, the application further outputs the result confidence.

[0202] The result confidence can be determined by the conductor point cloud integrity, the conductor segmentation probability, the single conductor clustering continuity, the catenary fitting residual error, the hanging point recognition quality and the historical ice-free benchmark matching quality: ;

[0203] Among them, representing conductor segmentation quality, representing single conductor clustering quality, representing catenary fitting quality, representing dropper identification quality, representing historical ice-free benchmark matching quality.

[0204] If the conductor or span outputs a low confidence label if the threshold is lower than a preset threshold; if the dropper is missing or the historical benchmark cannot be matched, an uncomputable label is output.

[0205] In another aspect, the present application also provides a point cloud-based ice thickness measurement system, comprising an acquisition module, a sag perception point cloud segmentation network and a processing module, and the point cloud-based ice thickness measurement system is used to execute any of the point cloud-based ice thickness measurement methods described above.

[0206] In another aspect, the present application also provides a terminal device, comprising a memory and a processor; the memory stores program code executable by the processor; and the program code is used to execute any of the point cloud-based ice thickness measurement methods described above.

[0207] For example, the program code can be divided into one or more modules / units, which are stored in the memory and executed by the processor to complete the present application. The one or more modules / units can be a series of computer program instruction segments capable of completing a specific function, which are used to describe the execution process of the program code in the terminal device.

[0211] As Figure 7 shown, Figure 7 is a schematic diagram of the hardware running environment of the terminal device related to the embodiments of the present application. The terminal device can be a desktop computer, a notebook computer, a palm computer, a cloud server and other computing devices. The terminal device can include, but is not limited to, a processor and a memory. Those skilled in the art can understand that the terminal device can also include input / output devices, network access devices, buses, etc.

[0209] The processor can be a central processing unit (CPU), and can also be other general-purpose processors, a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field-programmable gate array (FPGA) or other programmable logic device, discrete gate or transistor logic device, discrete hardware component, etc. The general-purpose processor can be a microprocessor or the processor can also be any conventional processor.

[0210] The memory can be an internal storage unit of the terminal device, such as a hard disk or a memory. The memory can also be an external storage device of the terminal device, such as a plug-in hard disk, a smart media card (SMC), a secure digital (SD) card, a flash card, etc. Further, the memory can also include both the internal storage unit and the external storage device of the terminal device. The memory is used to store the program code and other programs and data required by the terminal device. The memory can also be used to temporarily store data that has been output or will be output.

[0211] The terminal device and the ice thickness measurement system based on point cloud described above are created based on the ice thickness measurement method based on point cloud described above, and the technical effects and advantages thereof will not be repeated here. The technical features of the above-described embodiments can be combined in any manner. In order to make the description simple, all possible combinations of the technical features in the above-described embodiments are not described, however, as long as the combinations of the technical features do not exist contradictory, they should be considered as the scope of the present disclosure.

[0212] The above-described embodiments only express several embodiments of the present application, and the description is more specific and detailed, but it should not be understood as a limitation on the scope of the patent. It should be noted that for ordinary skilled in the art, without departing from the concept of the present application, several modifications and improvements can be made, which are all within the scope of the present application. Therefore, the scope of protection of the present application should be subject to the appended claims.

Claims

1. A method for measuring ice thickness based on point clouds, characterized in that, include: Acquire UAV laser point cloud data of the transmission line under test in its current state, and preprocess the UAV laser point cloud data to extract the initial point features of each point. The initial point features are input into the sag-sensing point cloud segmentation network for point-level semantic segmentation, outputting point-level semantic results for conductors, towers, and background. Based on the point-level semantic results, conductor point clouds and tower point clouds are extracted from the UAV laser point cloud data. The sag-sensing point cloud segmentation network includes a structure prior embedding module, an encoding module, a point-level decoding module, and a semantic prediction head, which are used to enhance the ability to distinguish between the slender continuous structure of conductors and the vertical structure of towers. Based on the tower point cloud and the tower spatial location, the span range between adjacent towers is determined, and within each span range, candidate conductor point clouds are extracted from the conductor point cloud. Single conductor separation is performed on the candidate conductor point cloud to obtain a single conductor point set. Based on the conductor category probability, local linearity, local point cloud density, and fitting residual confidence of each conductor point in the single conductor point set, the point-level fitting confidence of each conductor point is calculated, and a weighted catenary fitting is performed on the single conductor point set based on the point-level fitting confidence to obtain the three-dimensional centerline of the corresponding conductor. Identify the hanging points at both ends of the conductor based on the spatial adjacency relationship between the two ends of the three-dimensional centerline and the corresponding tower point cloud, and use the connecting line of the two hanging points as the reference chord to calculate the maximum vertical difference between the three-dimensional centerline and the reference chord to obtain the current sag. The sag increment is determined based on the current sag and the reference sag. The equivalent horizontal tension is calculated based on the reference sag, span length, and conductor foundation information under historical ice-free conditions. The additional load per unit length caused by icing is inverted based on the sag increment, span length, and equivalent horizontal tension. The unit length additional load is converted into unit length icing mass, and the equivalent icing thickness of the transmission line under test is calculated based on the bare conductor radius, icing density, and the circular equivalent model.

2. The method according to claim 1, characterized in that, The sag-aware point cloud segmentation network includes a structure prior embedding module, an encoding module, a point-level decoding module, and a semantic prediction head. The structure prior embedding module is used to describe the linearity, vertical distribution features and local density features of each point cloud in the local neighborhood, construct the structure prior features of each point cloud, and obtain the structure embedding features. The encoding module includes several encoding layers for performing layer-by-layer encoding processing on the structural embedding features. Each encoding layer includes a local geometric encoding branch, a directional continuous encoding branch, and a structural gated fusion unit. The local geometric encoding branch is used to extract spatial morphological differences in the neighborhood to obtain local geometric features. The directional continuous encoding branch is used to extract strip-level continuous features along the span direction to obtain directional continuous enhanced features. The structure-gated fusion unit is used to adaptively weighted fuse the local geometric features and the directional continuous enhancement features, and output the fused point features of each layer. The point-level decoding module is used to perform point-level semantic decoding on the fused point features of each layer, and to fuse the fused point features of each layer to obtain the decoded point-level features. The semantic prediction head is used to output the predicted probability of each point belonging to the conductor category or the tower category based on the decoded point-level features, and take the category corresponding to the highest probability as the semantic label of each point to generate the point-level semantic result.

3. The method according to claim 2, characterized in that, The steps for obtaining structural embedding features through the structural prior embedding module include: Based on the three-dimensional coordinates, laser reflection intensity, and neighborhood point cloud distribution of each point, the local point cloud density, local linear structure response, and vertical structure response of each point are calculated. The three-dimensional coordinates, laser reflection intensity, local point cloud density, local linear structure response, and vertical structure response are stitched together to obtain the enhanced input features; The enhanced input features are upscaled using a multilayer perceptron and mapped to a preset feature channel dimension to obtain the structural embedding features.

4. The method according to claim 2, characterized in that, The processing steps of the encoding module include: Through the local geometric coding branch, the farthest point sampling is performed on the point cloud corresponding to the structural embedding features of the current layer to obtain a low-resolution point set; A preset number of nearest neighbor sets are established for each sampling point in the low-resolution point set, and the local geometric features are output by combining the local geometric feature mapping function and the neighborhood max pooling operation. By using the directional continuous coding branches, the local span direction is determined according to the spatial location of the tower or the route, and the projected coordinates of each sampling point along the local span direction are calculated. The laser point cloud data is divided into directional strips according to the projection coordinates. The point features in each directional strip are pooled and aggregated to obtain strip-level features. The strip-level features of adjacent strips are propagated in one dimension along the strip arrangement direction. The propagated strip-level features are then back-propagated to the points in the corresponding strips to obtain the directional continuous enhancement features. The local geometric features and the directional continuous enhancement features are spliced ​​together along the feature channel dimension by the structure-gated fusion unit to obtain spliced ​​features; The stitched features are input into a gated multilayer perceptron and mapped to a preset interval using an activation function to determine the gate weights. The local geometric features and the directional continuous enhancement features are weighted by the gating weights, and the two weighting results are added element by element to obtain the fused point features of each layer.

5. The method according to claim 2, characterized in that, The steps for obtaining decoded point-level features through the point-level decoding module include: Obtain the input features of the current decoding layer, and use nearest neighbor interpolation to process the input features of the current decoding layer to obtain interpolated features, wherein the number of feature points of the interpolated features is restored to be consistent with the number of feature points of the fused feature points corresponding to the current decoding layer; The interpolated features are concatenated with the fused point features output from the corresponding encoding stage along the feature channel dimension, and then fused using a decoding multilayer perceptron to obtain the point-level features.

6. The method according to claim 1, characterized in that, The steps of extracting candidate conductor point clouds from the conductor point cloud within each span range, and performing single conductor separation on the candidate conductor point cloud to obtain a single conductor point set include: In the candidate point cloud of the conductor within each span, a spatial adjacency relationship is established based on the comparison result between the three-dimensional spatial distance between each point and the preset neighborhood radius threshold. Based on the spatial adjacency relationship, perform connectivity clustering on all candidate points of the conductor to determine the initial cluster. The principal direction of the initial cluster is calculated by principal component analysis, and the point cloud is divided into at least two slices along the principal direction. The distribution peaks of the point cloud in the horizontal and vertical directions within each slice are statistically analyzed. The point cloud at the corresponding slice position is divided into a preset number of sub-clusters along the gaps between the distribution peaks. The segmentation operation is performed iteratively until each sub-cluster has a single peak in both the horizontal and vertical distribution of the main direction. The final sub-cluster is determined as a single conductor point set corresponding to a spatially independent continuous conductor.

7. The method according to claim 1, characterized in that, The step of calculating the point-level fitting confidence of each traverse point based on the traverse class probability, local linearity, local point cloud density, and fitting residual confidence of each traverse point in the single traverse point set, and performing weighted catenary fitting on the single traverse point set based on the point-level fitting confidence to obtain the three-dimensional centerline of the corresponding traverse, further includes: Obtain the conductor class probability, local linearity, and local point cloud density of each conductor point in the single conductor point set; The confidence level of the fitting residuals is calculated based on the distance from the guide point to the initial fitted curve and the scaling parameter. The point-level fitting confidence of each traverse point is calculated by weighting the traverse category probability, local linearity, local point cloud density, and fitting residual confidence. The random sampling consensus algorithm is used to randomly sample and initially fit the points of the single conductor to obtain the set of interior points: In the local line coordinate system, a catenary model is used to describe the vertical coordinates of the conductor along the span direction, and a transverse model is used to describe the coordinates of the conductor along the transverse direction. Using the point-level fitting confidence as the weight, a weighted least squares fitting is performed on the set of interior points, and the catenary model parameters and transverse model parameters are solved to obtain the three-dimensional centerline of the conductor.

8. The method according to claim 1, characterized in that, The step of identifying the suspension points at both ends of the conductor based on the spatial adjacency relationship between the two ends of the three-dimensional centerline and the corresponding tower point cloud, and using the connecting line of the suspension points as a reference chord, calculating the maximum vertical difference between the three-dimensional centerline and the reference chord to obtain the current sag includes: Determine the spatial coordinates of the two hanging points, and use the line connecting the two hanging points as the reference chord. Calculate the vertical coordinates of the reference chord at each span direction; Obtain the vertical coordinates of the three-dimensional centerline at the same span direction, calculate the difference between the vertical coordinates of the reference chord and the vertical coordinates of the three-dimensional centerline, and take the maximum value of the difference between the two hanging points as the current sag.

9. The method according to claim 1, characterized in that, After the steps of converting the unit length additional load into unit length icing mass and calculating the equivalent icing thickness of the transmission line under test based on the bare conductor radius, icing density, and an equivalent ring model, the method further includes: The confidence level of the ice thickness measurement results is calculated based on the integrity of the conductor point cloud, the quality of conductor segmentation, the continuity of single conductor clustering, the three-dimensional centerline fitting residual, the quality of hanging point identification, and the quality of historical ice-free benchmark matching. When the confidence level of the result is lower than a preset threshold, a low confidence level flag is output. When the conductor hanging point is missing, the three-dimensional centerline fitting fails, or the historical ice-free benchmark cannot be matched, an uncalculated mark is output.

10. A point cloud-based ice thickness measurement system, characterized in that, The system is used to perform the point cloud-based ice thickness measurement method according to any one of claims 1 to 9; the point cloud-based ice thickness measurement system includes a point cloud acquisition and preprocessing module, a sag sensing point cloud segmentation module, a span division module, a single conductor separation module, a three-dimensional centerline fitting module, a hanging point identification and sag calculation module, and an ice thickness inversion module. The point cloud acquisition and preprocessing module is used to acquire UAV laser point cloud data of the transmission line under test in the current state, and to preprocess the UAV laser point cloud data to extract initial point features. The sag-aware point cloud segmentation module is used to perform point-level semantic segmentation on the initial point features based on the sag-aware point cloud segmentation network, to obtain point-level semantic results of conductors, towers and background, and to extract conductor point clouds and tower point clouds. The span division module is used to determine the span range between adjacent towers based on the tower point cloud and the tower spatial position. A single conductor separation module is used to extract candidate conductor point clouds from the conductor point cloud within each span range, and to separate single conductors from the candidate conductor point cloud to obtain a single conductor point set; The three-dimensional centerline fitting module is used to calculate the point-level fitting confidence based on the conductor category probability, local linearity, local point cloud density and fitting residual confidence, and to perform weighted catenary fitting on the single conductor point set based on the point-level fitting confidence to obtain the three-dimensional centerline of the corresponding conductor. The hanging point identification and sag calculation module is used to identify the hanging points at both ends of the conductor based on the spatial adjacency relationship between the two ends of the three-dimensional centerline and the corresponding tower point cloud, and to calculate the maximum vertical difference between the three-dimensional centerline and the reference chord line using the connecting line of the two hanging points as the reference chord line to obtain the current sag. The icing thickness inversion module is used to invert the additional load per unit length caused by icing based on the current sag, reference sag, span length, and historical ice-free conductor foundation information; convert the additional load per unit length into icing mass per unit length, and calculate the equivalent icing thickness of the transmission line under test based on the bare conductor radius, icing density, and circular equivalent model.

Images