Complex scene strain tower jumper comprehensive analysis system

Abstract

The application discloses a complex scene tension tower jumper wire comprehensive analysis system based on the technical field of data processing, and belongs to the technical field of data processing, and comprises a multi-source data receiving module, which receives visible light, infrared thermal imaging and laser point cloud data collected by a unmanned aerial vehicle; a multi-modal feature extraction module, which extracts jumper wire surface texture, temperature gradient and spatial geometric features; an icing voxel construction module, which generates a three-dimensional voxel matrix representing the distribution of icing through a cross-modal cross-attention mechanism; a twin load mapping module, which maps the three-dimensional voxel matrix to a jumper wire catenary digital twin grid, and calculates a non-uniform load vector according to the icing voxel volume of each grid unit; and a shape stress solving module, which updates node coordinates and stress through nonlinear iteration solving, and outputs jumper wire shape and suspension string inclination angle data. The application converts sensing data into non-uniform load boundary conditions bound with physical grids, replaces the uniform static load assumption, and reduces the calculation error of the stress and shape of the jumper wire under non-uniform icing.

Description

Technical Field

[0001] This invention belongs to the field of data processing technology, specifically relating to a comprehensive analysis system for tension tower jumpers in complex scenarios. Background Technology

[0002] In the field of transmission line icing monitoring, assessing the stress state of tension tower jumpers is fundamental to ensuring the safe operation of the power grid. Current technologies typically employ drones equipped with one or more sensors to acquire jumper images and point cloud data, then manually input design ice thickness parameters or estimate the equivalent ice thickness based on visual algorithms. In subsequent mechanical calculations, existing methods apply the estimated ice thickness as a uniform icing thickness to the physical model of the entire jumper span. This approach treats icing as a static load uniformly distributed along the line direction and solves for jumper sag variation, stress distribution, and suspension string skew based on this uniform load boundary conditions.

[0003] The aforementioned existing technology has a core flaw in the icing load application process. Under complex weather conditions, the icing morphology on the windward and leeward sides of the jumper differs, and the icing distribution on both sides of the tension string also exhibits spatial heterogeneity. Simplifying the actual non-uniform spatial distribution of icing into a static load uniformly distributed along the span will cause the load boundary conditions input to the mechanical calculation model to deviate from the actual stress state of the jumper, thus resulting in discrepancies between the jumper morphology evolution and the internal stress solution results. Summary of the Invention

[0004] The purpose of this invention is to provide a comprehensive analysis system for tension tower jumpers in complex scenarios. This system receives visible light images, infrared thermal images, and laser point cloud data collected by UAVs, and extracts texture features, temperature gradient features, and spatial geometric features of the jumper surface. Based on the extracted features, a three-dimensional voxel matrix representing the icing distribution is generated through a cross-modal cross-attention mechanism. This matrix is ​​mapped to a digital twin mesh of the jumper catenary, and a non-uniformly distributed load vector is calculated according to the voxel volume of each mesh unit. The node coordinates and stress are updated through nonlinear iterative solution, and the jumper morphology evolution results and skew angle data are output, thereby effectively solving the problems existing in the prior art.

[0005] To achieve the above objectives, the present invention adopts the following technical solution:

[0006] The comprehensive analysis system for tension tower jumpers in complex scenarios includes: a multi-source data receiving module, used to receive visible light images, infrared thermal images, and laser point cloud data of tension tower jumpers collected by UAVs;

[0007] The multimodal feature extraction module is used to extract jumper surface texture features, temperature gradient features, and spatial geometric features from the visible light image, the infrared thermal image, and the laser point cloud data, respectively, using a multimodal feature extraction network.

[0008] The ice-covered voxel construction module is used to generate a three-dimensional voxel matrix representing the ice-covered distribution based on the extracted jumper surface texture features, temperature gradient features, and spatial geometric features through a cross-modal cross attention mechanism.

[0009] The twin load mapping module is used to map the three-dimensional voxel matrix to a pre-established jumper catenary digital twin mesh, and calculate the non-uniformly distributed load vector according to the ice-covered voxel volume corresponding to each grid cell in the jumper catenary digital twin mesh.

[0010] The morphological stress solution module is used to input the non-uniformly distributed load vector into the nonlinear iterative solver, update the node coordinates and element stress of the jumper catenary digital twin mesh, and output the jumper morphological evolution results and suspension string deflection angle data under icing conditions.

[0011] Furthermore, the multimodal feature extraction network includes a visible light convolution branch, an infrared convolution branch, and a point cloud map convolution branch;

[0012] The visible light convolution branch performs multi-scale convolution pooling operations on the visible light image, and outputs a texture feature map that fuses local texture and global contour.

[0013] The infrared convolution branch performs thermal radiation gradient difference calculation on the infrared thermal image to extract the temperature difference boundary features between the iced area and the bare line area as temperature gradient features.

[0014] The point cloud graph convolutional branch constructs the topological adjacency graph of the laser point cloud data, and aggregates the three-dimensional coordinate vectors and reflection intensity vectors of neighboring points through the graph convolution kernel, outputting spatial geometric features including jumper three-dimensional spatial skeleton information;

[0015] The texture feature map, the temperature gradient feature, and the spatial geometric feature are projected onto a feature space of a unified dimension.

[0016] Furthermore, the ice-covered voxel construction module implements the cross-modal cross-attention mechanism through a cross-modal cross-attention module in the following manner:

[0017] The feature sequence of each spatial location point in the spatial geometric features is used as the query vector, and the multimodal joint feature sequence after the texture feature map and the temperature gradient feature are concatenated as the key vector and value vector.

[0018] By calculating the multi-head self-attention similarity matrix between the query vector and the key vector, the attention weight coefficients of each spatial location point feature sequence in the spatial geometric features to each feature element in the multimodal joint feature sequence are determined.

[0019] The value vector is weighted and summed using the attention weight coefficients to filter out texture features and spatial misalignment interference caused by the drone's shooting perspective shift, thereby generating the three-dimensional voxel matrix aligned with three-dimensional spatial coordinates and representing ice thickness and density.

[0020] Furthermore, the digital twin mesh of the jumper catenary is discretized according to the actual catenary equation of the tension tower jumper, dividing the jumper into multiple mesh units with definite spatial topological connections.

[0021] The twin load mapping module uses a spatial bounding box collision detection algorithm to cluster and statistically analyze the voxels in the three-dimensional voxel matrix that fall within the spatial boundaries of each grid cell.

[0022] For each grid cell, the local icing mass of the corresponding grid cell is calculated by multiplying the voxel volume obtained from clustering with the preset icing density parameter.

[0023] Based on the gravitational acceleration vector and the normal vector of the grid cell, the local icing mass is converted into a non-uniformly distributed load vector along the gravity direction and perpendicular to the tangent direction of the grid cell, and the non-uniformly distributed load vector is applied to the nodes of the corresponding grid cell.

[0024] Furthermore, the nonlinear iterative solver uses the Newton-Raphson method for solving the problem;

[0025] In each iteration step, the tangent stiffness matrix of the jumper catenary digital twin mesh is constructed based on the node coordinates of the current iteration step;

[0026] The nodal unbalanced force vector after applying the non-uniformly distributed load vector is multiplied by the inverse of the tangent stiffness matrix to obtain the nodal displacement increment.

[0027] The node displacement increment is superimposed on the node coordinates of the current iteration step, and the strain tensor and stress tensor of each mesh element are updated according to the topological connection relationship of the jumper catenary digital twin mesh.

[0028] When the norm of the unbalanced force vector at the node is less than a preset convergence threshold, the iteration is terminated and the node coordinates and element stress in the final stable state are output as the result of the jumper morphology evolution.

[0029] Furthermore, when calculating the skew angle data of the suspension string, the morphological stress solving module extracts the element stress at the end node connected to the suspension string hardware in the digital twin mesh of the jumper catenary.

[0030] The unit stress at the end node is decomposed into a tension component along the jumper tangent and a lateral force component perpendicular to the jumper tangent.

[0031] The spatial resultant tension vector of the suspension string is calculated based on the vector sum of its own gravity vector, the tension component, and the lateral force component.

[0032] Calculate the spatial angle between the spatial resultant tension vector and the vertically downward gravity direction, and project the spatial angle onto the cross section and longitudinal section of the transmission line to generate the suspension string deflection angle data, which includes the deflection angle along the line direction and the deflection angle across the line direction.

[0033] Preferably, the ice-covered voxel construction module implements the cross-modal cross-attention mechanism through a cross-modal cross-attention module in the following manner:

[0034] The cross-modal attention module uses the feature sequence of each spatial location point in the spatial geometric features as the key vector and the value vector, and uses the multimodal joint feature sequence after concatenating the texture feature map and the temperature gradient feature as the query vector.

[0035] By calculating the multi-head self-attention similarity matrix between the query vector and the key vector, the attention weight coefficient of each feature element in the multimodal joint feature sequence to the spatial location point feature sequence is determined.

[0036] The value vector is weighted and summed using the attention weight coefficients to filter out texture features and spatial misalignment interference caused by the drone's shooting perspective shift, thereby generating the three-dimensional voxel matrix aligned with three-dimensional spatial coordinates and representing the ice thickness and density.

[0037] Preferably, the cross-modal cross-attention module introduces a modality credibility evaluation mechanism before calculating the multi-head self-attention similarity matrix;

[0038] For the texture feature components in the multimodal joint feature sequence, calculate the texture complexity entropy value of the corresponding pixel region. When the texture complexity entropy value is lower than the preset smoothing threshold, it is determined that the visible light features are degraded due to the light transmittance of the ice layer, and the proportion of the texture feature components in the query vector is reduced.

[0039] For the temperature gradient feature components, the thermal radiation gradient variance is calculated. When the thermal radiation gradient variance is lower than a preset uniformity threshold, it is determined that the infrared features have degraded due to the homogenization of ambient temperature, and the proportion of the temperature gradient feature components in the query vector is reduced.

[0040] The subsequent weighted summation calculation is performed based on the adjusted query vector to suppress the interference of low-quality modal features on the three-dimensional voxel matrix.

[0041] Preferably, the twin load mapping module deploys a voxel morphology filter before executing the spatial bounding box collision detection algorithm;

[0042] The voxel morphology filter performs erosion and dilation operations on the three-dimensional voxel matrix in sequence to remove discrete isolated noise voxels attached to the surface of the three-dimensional voxel matrix.

[0043] When calculating the local icing quality, for each grid cell, the number of connected components in the voxel cluster after morphological filtering is counted.

[0044] When the number of connected components is greater than one, retain the largest main connected component voxel cluster and remove the dangling non-connected voxel clusters.

[0045] The volume of the voxel cluster in the main connected domain is used to replace the original voxel volume obtained by clustering and participates in the calculation of the local icing quality, thereby eliminating abnormal fluctuations in the calculation of icing volume caused by fluctuations in laser point cloud density.

[0046] Preferably, the nonlinear iterative solver monitors the eigenvalues ​​of the deformation gradient matrix of each mesh element in real time during each iteration step;

[0047] When the ratio of the maximum eigenvalue to the minimum eigenvalue of the deformation gradient matrix of any mesh cell exceeds a preset distortion threshold, the mesh cell is determined to have mesh distortion.

[0048] For the mesh cell that has undergone mesh distortion, a local mesh reconstruction strategy is triggered. Based on the adjacent nodes of the mesh cell, new nodes are inserted and the distorted mesh cell is divided into multiple sub-mesh cells.

[0049] The non-uniformly distributed load vector of the original distorted mesh element is distributed to each of the sub-mesh elements according to the principle of volume equivalence, and the node coordinates of the updated jumper catenary digital twin mesh are smoothed using the least squares method to prevent iterative divergence caused by large local deformation.

[0050] Preferably, the morphological stress solving module simultaneously acquires the wind field streamline data of the micro-topography area where the tension tower is located during the calculation of the suspension string deflection angle data.

[0051] Based on the wind field streamline data, the aerodynamic drag vector of each grid cell is calculated, and the aerodynamic drag vector is superimposed on the non-uniformly distributed load vector to form a comprehensive dynamic load vector.

[0052] The Newton-Raphson method iteration is re-executed using the comprehensive dynamic load vector to extract the spatial resultant tension vector change trajectory under dynamic time sequence.

[0053] The extreme points in the trajectory of the change of the spatial resultant tension vector are extracted, and a deflection angle envelope space containing the maximum longitudinal deflection angle and the maximum transverse deflection angle is constructed. The deflection angle envelope space is used as a component of the output of the deflection angle data of the suspension string to characterize the maximum range of motion of the suspension string under wind-ice coupling conditions.

[0054] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0055] 1) Accurate characterization of non-uniformly distributed icing load based on cross-modal attention fusion: This invention uses a cross-modal cross-attention module to perform cross-modal weight allocation on spatial geometric features, texture features, and temperature gradient features, generating a three-dimensional voxel matrix aligned with three-dimensional spatial coordinates. The three-dimensional voxel matrix is ​​then mapped to the digital twin mesh of the jumper catenary. The non-uniformly distributed load vector is calculated according to the icing voxel volume corresponding to each mesh unit. This technique transforms multi-source heterogeneous sensor data into non-uniformly distributed load boundary conditions bound to the physical mesh, replacing the static load assumption of uniform distribution along the span in the prior art. It corrects the input deviation of the mechanical boundary conditions, allowing the spatial heterogeneity of icing to be fully reflected in the mechanical calculation, thereby reducing the calculation error of the stress state and morphological evolution process of the jumper under non-uniform icing conditions.

[0056] 2) To address the issue of fluctuating quality in multi-source sensor data, a multi-level robustness enhancement mechanism is constructed at key stages to improve system stability in complex field environments: In the multi-modal feature fusion stage, a modal reliability assessment mechanism is introduced. By calculating the texture complexity entropy and thermal radiation gradient variance, the proportion of degraded modal features in the query vector is reduced, suppressing the interference of low-quality sensor data on the generation of the 3D voxel matrix; Before load mapping, a voxel morphological filter is deployed to remove isolated noise voxels and retain the voxel clusters of the main connected domain, eliminating the impact of point cloud density fluctuations on the calculation of local icing quality; In the solution stage, by monitoring the eigenvalues ​​of the deformation gradient matrix of the grid cells in real time and triggering a local grid reconstruction strategy, iterative divergence caused by large local deformations is avoided.

[0057] 3) Expand the ability to analyze wind-ice coupling conditions and output the dynamic motion range envelope of the suspension string: Combine wind field streamline data to calculate the aerodynamic drag vector and superimpose it into the load to form a comprehensive dynamic load. This enables the output suspension string skew angle envelope space to reflect the maximum motion range under wind-ice coupling conditions, providing more comprehensive data support for the safety verification of transmission lines under extreme weather conditions. Attached Figure Description

[0058] Figure 1 This is a schematic diagram illustrating the structure and data flow of the comprehensive analysis system for tension tower jumpers in complex scenarios based on the present invention.

[0059] Figure 2This is a flowchart of the multimodal feature extraction and modal reliability evaluation process of the present invention;

[0060] Figure 3 This is a flowchart of the cross-modal cross-attention mechanism and voxel morphological filtering of the present invention;

[0061] Figure 4 This is a flowchart of the non-uniformly distributed load vector calculation of the present invention;

[0062] Figure 5 This is a flowchart of the nonlinear iterative solution and mesh distortion processing of the present invention;

[0063] Figure 6 This is a flowchart of the analysis process for the coupling of the suspension string deflection angle and wind and ice in this invention. Detailed Implementation

[0064] To make the technical solution of the present invention clearer, the technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0065] In one specific embodiment, such as Figure 1 As shown, the comprehensive analysis system for tension tower jumpers in complex scenarios provided by the present invention includes: a multi-source data receiving module, a multi-modal feature extraction module, an ice-covering voxel construction module, a twin load mapping module, and a morphological stress solving module.

[0066] The multi-source data receiving module is used to receive visible light images, infrared thermal images, and laser point cloud data of tension tower jumpers collected by UAVs, specifically:

[0067] After receiving data, the multi-source data receiving module performs a spatiotemporal synchronization and alignment operation to unify visible light images, infrared thermal images, and laser point cloud data under the same world coordinate system and time reference. The spatiotemporal synchronization and alignment operation uses the pose data output by the UAV's integrated navigation system, combined with the intrinsic and extrinsic parameter calibration results of each sensor, to perform coordinate transformation and timestamp matching on the data collected by different sensors. During the coordinate transformation process, based on the rotation matrix and translation vector obtained from the sensor extrinsic parameter calibration, the local coordinate system data of each sensor is transformed to the UAV's body coordinate system. Then, based on the pose data output by the integrated navigation system, the body coordinate system data is transformed to the geodetic coordinate system, so that the jump line target at the same spatial location has a one-to-one mapping relationship in different modal data. During the timestamp matching process, the acquisition timestamp of the laser point cloud data is used as the reference, and visible light images and infrared thermal images with timestamp differences within a preset time window are selected to complete the time synchronization of multi-source data.

[0068] The multimodal feature extraction module utilizes a multimodal feature extraction network to extract surface texture features, temperature gradient features, and spatial geometric features from visible light images, infrared thermal images, and laser point cloud data, respectively. Specifically:

[0069] Texture feature extraction: The multimodal feature extraction network performs convolution operations on the input visible light image to extract the surface texture information of the jumper wire, including the gray-level changes and edge contours of the icy and bare wire areas, and generates surface texture features of the jumper wire. The convolution operation adopts a multi-layer cascaded convolution structure, extracting texture information at different scales through convolution kernels of different sizes. Shallow convolution kernels extract subtle texture changes on the surface of the jumper wire, while deep convolution kernels extract the overall contour features of the jumper wire.

[0070] Temperature gradient feature extraction: The multimodal feature extraction network performs gradient difference operation on the input infrared thermal image to extract the thermal radiation value change information of different regions and generate temperature gradient features that characterize the temperature difference distribution between the iced area and the bare line area; the gradient difference operation calculates the difference in thermal radiation value between adjacent pixels to obtain the spatial rate of change of thermal radiation value, thereby locating the temperature difference boundary between the iced area and the bare line area.

[0071] Spatial geometric feature extraction: The multimodal feature extraction network performs spatial topology construction and feature aggregation on the input laser point cloud data to extract the spatial geometric information of the jumper, including three-dimensional spatial coordinates and contour morphology information, generating spatial geometric features including the spatial location and shape of the jumper. Among them, spatial topology construction establishes neighborhood relationships for each point cloud data point to realize the structured representation of the point cloud data; feature aggregation extracts the global spatial morphological features of the jumper by aggregating the coordinates and reflection intensity information of neighboring points.

[0072] After extraction, the multimodal feature extraction module maps the extracted texture features, temperature gradient features, and spatial geometric features to a feature space of a unified dimension and outputs them to the ice-covered voxel construction module.

[0073] The ice-covering voxel construction module, through a cross-modal cross-attention module, uses spatial geometric features as query vectors to weight texture and temperature features, generating a three-dimensional voxel matrix representing ice thickness and density. Specifically:

[0074] The cross-modal attention module converts spatial geometric features into a query vector sequence, and concatenates texture features and temperature gradient features into a key vector sequence and a value vector sequence. It calculates the similarity between the query vector and the key vector through vector dot product operation. The magnitude of the dot product result represents the degree of correlation between the two feature vectors. The higher the degree of correlation, the greater the corresponding attention weight. The attention weight is used to perform a weighted summation of the value vectors to generate a fusion feature aligned with the three-dimensional spatial coordinates.

[0075] Subsequently, based on the fusion features, the ice-covered voxel construction module performs uniform meshing on the three-dimensional space where the jumper is located, generating a three-dimensional voxel mesh. Each voxel mesh corresponds to a cubic region in the three-dimensional space and has a defined spatial coordinate range. Each voxel mesh corresponds to a fusion feature vector. The numerical value of the fusion feature vector represents the ice thickness and density parameters at the corresponding voxel position. Finally, a three-dimensional voxel matrix is ​​generated, realizing the three-dimensional digital representation of the ice distribution on the jumper surface.

[0076] The twin load mapping module maps the three-dimensional voxel matrix to a pre-established jumper catenary digital twin mesh, and calculates the non-uniformly distributed load vector according to the ice-covered voxel volume corresponding to each mesh cell in the jumper catenary digital twin mesh, specifically:

[0077] The digital twin mesh of the jumper catenary is pre-constructed based on the design parameters of the jumper and the mechanical properties of the catenary. The initial catenary shape of the jumper is calculated by the following catenary equation:

[0078] ,

[0079] in, For jump lines on the x-axis The vertical coordinates at that location, The horizontal tension of the jumper wire, The mass per unit length of the jumper wire. It is the acceleration due to gravity. Let x be the x-coordinate of the lowest point of the catenary. Here are the vertical coordinates of the lowest point of the catenary. It is a hyperbolic cosine function;

[0080] Based on the initial shape of the jumper obtained from the above catenary equation, the continuous catenary structure is divided into multiple mesh elements with definite spatial topological connections using the finite element discretization method. Adjacent mesh elements are connected through common nodes, and all mesh elements and nodes together constitute the digital twin mesh of the jumper catenary. Each mesh element corresponds to a unique spatial location interval and has definite material property parameters, including element length, cross-sectional area, elastic modulus, and Poisson's ratio.

[0081] The twin load mapping module establishes the spatial coordinate mapping relationship between the three-dimensional voxel matrix and the jumper catenary digital twin mesh. For each mesh cell, it counts the number and total volume of voxels falling into the spatial boundary of the mesh cell. Based on the total volume of voxels and the preset icing density parameter, it calculates the local icing mass corresponding to the mesh cell and then converts the local icing mass into a non-uniformly distributed load vector applied to the mesh cell nodes.

[0082] The morphological stress solution module inputs the non-uniformly distributed load vector into the nonlinear iterative solver, updates the nodal coordinates and element stresses of the jumper catenary digital twin mesh, and outputs the jumper morphological evolution results and suspension string deflection angle data under icing conditions, specifically:

[0083] The nonlinear iterative solver is based on the input non-uniformly distributed load vector. Taking the initial node coordinates of the jumper catenary digital twin mesh as the starting point of the iteration, it calculates the mesh stiffness matrix and nodal unbalanced forces in each iteration step, solves the nodal displacement increment and updates the nodal coordinates, and calculates the strain and stress of each mesh element until the iteration results meet the preset convergence conditions. Among them, the nodal unbalanced force is the difference between the external force and the internal force on the node. During the iteration process, the nodal unbalanced force gradually decreases. When its norm is less than the preset convergence threshold, the iteration is determined to be converged.

[0084] After iterative convergence, the morphological stress solution module outputs the final grid node coordinates as the jumper morphology evolution result. The change in node coordinates characterizes the spatial morphological change of the jumper under the action of icing load. At the same time, based on the stress distribution of the grid end nodes, the sag string deflection angle data is calculated.

[0085] In this embodiment, the various features output by the multimodal feature extraction module need to be projected onto a feature space of a unified dimension to meet the input requirements of the cross-modal attention module. The dimensional correspondence of various features is shown in Table 1.

[0086] Table 1

[0087]

[0088] In Table 1, all types of single-modal features are mapped to a unified 256-dimensional feature space. This ensures that when the cross-modal attention module assigns weights, each feature vector has the same dimensionality matching, avoiding feature fusion bias caused by dimensionality differences. Specifically, the dimensionality mapping of surface texture features and temperature gradient features is achieved through 1×1 convolution operations, while the dimensionality mapping of spatial geometric features is achieved through fully connected layer operations. All mapping operations preserve the core physical meaning of the features and do not change the spatial location relationships corresponding to the features.

[0089] This embodiment generates a three-dimensional voxel matrix representing the ice distribution by synchronous alignment of multi-source data and extraction of multi-modal features, combined with cross-modal feature fusion. The three-dimensional voxel matrix is ​​then mapped onto the digital twin mesh of the jumper catenary to generate a non-uniformly distributed load vector. The morphological evolution of the jumper and the sag string deflection angle data are obtained through nonlinear iterative solution, realizing a complete characterization of the spatial heterogeneity of ice in the mechanical analysis of the jumper and correcting the mechanical boundary condition deviation caused by the uniform load assumption.

[0090] In a preferred embodiment, such as Figure 2As shown, the multimodal feature extraction network includes visible light convolution branches, infrared convolution branches, and point cloud map convolution branches.

[0091] The visible light convolution branch performs multi-scale convolution and pooling operations on the visible light image, outputting a texture feature map that fuses local texture and global contour, specifically:

[0092] The visible light convolution branch adopts a multi-scale convolution architecture, which includes multiple cascaded convolutional units and pooling units. Each convolutional unit contains two 3×3 convolutional layers and one ReLU activation layer. The stride of the convolutional layers is set to 1, and the padding is set to 1 to ensure that the size of the feature map remains consistent before and after the convolution operation. The pooling unit adopts a 2×2 max pooling operation with a stride of 2 to downsample the feature map and expand the receptive field of subsequent convolutional layers.

[0093] The visible light convolution branch extracts features from different receptive fields in convolutional units at different levels. Shallow convolutional units extract local texture details of the jumper surface, including information such as surface roughness of icy areas and ice edge. Deep convolutional units extract global contour information of the jumper, including information such as the overall direction and sag shape of the jumper. The visible light convolution branch upsamples and fused the convolutional features at different levels through a feature pyramid structure. Deep features are restored to the same size as shallow features through upsampling operations, and then concatenated with shallow features in the channel dimension. The final output is a texture feature map that contains both local texture and global contour.

[0094] The infrared convolution branch performs thermal radiation gradient difference calculation on the infrared thermal image, extracting the temperature difference boundary features between the iced area and the bare line area as the temperature gradient features, specifically:

[0095] Normalization processing: The infrared convolution branch first performs pixel value normalization processing on the input infrared thermal image. The thermal radiation value of the infrared thermal image is converted into a standardized value in the range of 0 to 1 through linear transformation. The minimum value of the thermal radiation value is mapped to 0 and the maximum value is mapped to 1.

[0096] Gradient difference calculation: The Sobel gradient operator is used to perform horizontal and vertical gradient difference calculation on the normalized infrared thermal image. The Sobel operator includes a horizontal gradient kernel and a vertical gradient kernel, which are convolved with the infrared thermal image to obtain the gradient components of the thermal radiation value in the horizontal and vertical directions.

[0097] Gradient magnitude and direction calculation: The thermal radiation gradient magnitude and gradient direction of each pixel are calculated based on the gradient components; the gradient magnitude is the square root of the sum of the squares of the horizontal gradient components and the vertical gradient components, and the gradient direction is the arctangent of the ratio of the vertical gradient component to the horizontal gradient component.

[0098] Temperature difference boundary extraction: Thresholding is performed on the gradient magnitude to retain pixel areas with gradient magnitudes greater than the preset gradient threshold. These areas are the temperature difference boundary between the icing area and the bare wire area.

[0099] Feature encoding: The gradient distribution information of the temperature difference boundary region is converted into a fixed-dimensional feature vector through a fully connected layer, generating temperature gradient features that include the location, magnitude and direction of the temperature difference boundary.

[0100] The point cloud graph convolutional branch constructs a topological adjacency graph of the laser point cloud data. The graph convolution kernel aggregates the 3D coordinate vectors and reflection intensity vectors of neighboring points, outputting spatial geometric features including the 3D spatial skeleton information of the jumper lines. Specifically:

[0101] Point cloud preprocessing: The convolutional branch of the point cloud image first preprocesses the input laser point cloud data, that is, it uses a statistical filtering algorithm to calculate the average distance between each point and its k nearest neighbors, removes outliers whose average distance is greater than a preset distance threshold, and filters out noise points to obtain the filtered jumper point cloud data.

[0102] Construction of topological adjacency graph: Based on the k-nearest neighbor algorithm, a topological adjacency relationship is constructed for each point cloud data point. The connection between each data point and its k neighboring points constitutes the edge of the topological adjacency graph. All data points and adjacent edges together constitute the topological adjacency graph of the laser point cloud data. At this time, each data point is regarded as a node in the graph structure.

[0103] Graph convolution operation: The 3D coordinate vector of each node is concatenated with the reflection intensity vector to form the initial feature vector of that node. A graph convolution kernel is then used to perform a convolution operation on the topological adjacency graph, aggregating the feature information of each node and its neighboring nodes. The graph convolution operation is implemented using the following formula:

[0104] ,

[0105] in, For the first In layer graph convolution, the first The output feature vector of each node For the first The set of neighboring nodes of a node. and The first The node and the first The degree of each node. For the first The learnable weight matrix of layer graph convolution. For the first Bias vector of layer graph convolution It is a non-linear activation function. For the first In layer graph convolution, the first The input feature vectors of the neighboring nodes;

[0106] Global feature aggregation: Through multi-layer graph convolution operations, a larger range of neighborhood features are aggregated step by step to extract the global spatial structure information of the jumper point cloud. Finally, through global average pooling, the features of all nodes are aggregated into a fixed-dimensional global feature vector, and the output is spatial geometric features including the jumper 3D spatial skeleton information.

[0107] Furthermore, the texture feature map, temperature gradient feature, and spatial geometric feature are projected onto a feature space of a unified dimension for use by the cross-modal attention module, specifically:

[0108] Texture feature map: Through global average pooling and 1×1 convolution operation, the two-dimensional feature map is converted into a one-dimensional feature vector and mapped to a preset unified feature dimension;

[0109] Temperature gradient features: The gradient distribution encoding vector is mapped to a unified feature dimension through fully connected layer operations;

[0110] Spatial geometric features: Through global pooling and fully connected layer operations, the node feature sequence output by graph convolution is aggregated into a global feature vector and mapped to a unified feature dimension.

[0111] The icing voxel module implements the cross-modal cross-attention mechanism through the cross-modal cross-attention module in the following manner:

[0112] The spatial geometric features are used as key and value vectors, and the multimodal joint feature sequence resulting from concatenating the texture feature map and temperature gradient features is used as the query vector. Specifically, the spatial geometric features are feature sequences containing multiple spatial points, each feature vector corresponding to a specific coordinate position in three-dimensional space. The cross-modal attention module performs linear transformations on these feature sequences to generate key vector matrices. AND value vector matrix Simultaneously, the cross-modal attention module concatenates the texture feature map and temperature gradient feature along the feature dimension to generate a multimodal joint feature sequence. A linear transformation is then performed on this multimodal joint feature sequence to generate a query vector matrix. ;

[0113] A modal credibility assessment mechanism is introduced, specifically as follows:

[0114] For the texture feature components in the multimodal joint feature sequence, the texture complexity entropy value of the corresponding pixel region is calculated. The texture complexity entropy value is calculated using the following formula:

[0115] ,

[0116] in, This represents the texture complexity entropy value. This represents the total number of pixel gray levels. The grayscale value within the corresponding pixel area The probability of a pixel appearing;

[0117] When the texture complexity entropy value is lower than the preset smoothing threshold, it is determined that the visible light features have degraded due to the translucency of the ice layer, and the proportion of texture feature components in the query vector is reduced. Specifically: for each pixel block corresponding to the jumper area in the visible light image, the gray-level distribution probability of the pixel block is calculated, and the texture complexity entropy value is calculated based on the above formula. The lower the texture complexity entropy value, the smoother the texture of the pixel block, and the visible light image cannot effectively distinguish between the icy area and the bare line area, that is, the visible light features have degraded; when the texture complexity entropy value is lower than the preset smoothing threshold, the weight proportion of the texture feature component corresponding to the pixel block in the multimodal joint feature sequence is reduced according to the preset attenuation coefficient.

[0118] For the characteristic components of the temperature gradient, the variance of the thermal radiation gradient is calculated using the following formula:

[0119] ,

[0120] in, The variance of the thermal radiation gradient. This represents the total number of pixels in the corresponding area of ​​the infrared thermal image. For the first The thermal radiation gradient magnitude of each pixel This represents the average value of the thermal radiation gradient amplitude of all pixels within the corresponding region.

[0121] When the thermal radiation gradient variance is lower than a preset uniformity threshold, it is determined that the infrared features have degraded due to the homogenization of ambient temperature. The proportion of the temperature gradient feature component in the query vector is reduced. Specifically, for each pixel block in the corresponding jumper area of ​​the infrared thermal image, the average value of the thermal radiation gradient amplitude of all pixels in the pixel block is calculated. The thermal radiation gradient variance is calculated based on the above formula. The lower the thermal radiation gradient variance, the more uniform the thermal radiation gradient distribution in the pixel block. The infrared thermal image cannot effectively distinguish the temperature difference boundary between the iced area and the bare line area, that is, the infrared features have degraded. When the thermal radiation gradient variance is lower than a preset uniformity threshold, the weight proportion of the temperature gradient feature component corresponding to the pixel block in the multimodal joint feature sequence is reduced according to a preset attenuation coefficient.

[0122] Finally, the adjusted multimodal feature sequence is output, which is the adjusted query vector;

[0123] By calculating the multi-head self-attention similarity matrix between the adjusted query vector and the key vector, the attention weight coefficients of each feature element in the multimodal joint feature sequence to the spatial location point feature sequence are determined; the core calculation process of multi-head self-attention is implemented through the following formula:

[0124] ,

[0125] in, To query the vector matrix, The key vector matrix, It is a value vector matrix. Let be the dimension of the key vector, and ⊤ be the matrix transpose operation. It is a normalized exponential function;

[0126] The calculation of multi-head self-attention involves performing h linear projections on the query vector, key vector, and value vector respectively, calculating h sets of attention outputs, and then concatenating and linearly transforming the h sets of outputs to obtain the final multi-head attention output, ensuring that feature association information of different dimensions can be captured. Based on the above formula, the dot product of the query vector matrix and the key vector matrix is ​​calculated, scaled by dividing by the square root of the key vector dimension, and then normalized by the softmax function to obtain the attention weight coefficient matrix. Each element in the matrix corresponds to the attention weight of a feature element in the multimodal joint feature sequence to a feature vector in the spatial location point feature sequence.

[0127] The attention weight coefficients are used to perform a weighted summation on the value vectors to filter out the interference of texture features and spatial position misalignment caused by the perspective shift of the drone shooting, and generate a three-dimensional voxel matrix of ice thickness and density aligned with three-dimensional spatial coordinates. Specifically, the attention weight coefficient matrix and the value vector matrix are multiplied to obtain a weighted fused feature sequence. Each feature vector in the feature sequence corresponds to a certain coordinate position in three-dimensional space, realizing the spatial alignment of texture features, temperature gradient features and spatial geometric features.

[0128] Subsequently, the ice-covered voxel module maps the weighted fused feature sequence into a three-dimensional voxel grid. Each voxel grid corresponds to a fused feature vector. The two dimensions of the fused feature vector represent the ice thickness and ice density at the corresponding voxel position, respectively, and finally generate a three-dimensional voxel matrix.

[0129] In this embodiment, after the weight adjustment is completed through the modality credibility assessment mechanism, an updated multimodal joint feature sequence is generated. A new query vector matrix is ​​generated based on the updated multimodal joint feature sequence, and then subsequent multi-head self-attention calculation and feature weighted summation operations are performed to ensure that the features of degraded modalities do not negatively affect the accuracy of fused features. Table 2 shows the threshold setting and weight adjustment rules of the modality credibility assessment mechanism. These rules are used to quantify the modality feature weight adjustment range under different degrees of degradation to ensure that the interference of low-quality modalities is effectively suppressed during feature fusion.

[0130] Table 2

[0131]

[0132] In Table 2, the smoothing threshold and uniformity threshold are calibrated based on historical data from on-site inspections of transmission lines. The weight adjustment coefficient is used to directly multiply the feature components of the corresponding mode, thereby dynamically adjusting the weights of degraded mode features. Specifically, for severely degraded mode features, the weight adjustment coefficient is set to 0.2, retaining only a small amount of feature information for auxiliary fusion, thus avoiding significant interference from low-quality features in the generation of the three-dimensional voxel matrix.

[0133] This embodiment uses three independent convolutional branches to extract features from different modalities, ensuring that the core feature information of each modality is effectively extracted. It also uses a cross-modal attention module to achieve spatial alignment and fusion of multimodal features. Furthermore, it introduces a modality credibility evaluation mechanism to dynamically adjust the feature weights of degenerate modalities, effectively suppressing the interference of low-quality sensor data on the generation of the three-dimensional voxel matrix and improving the accuracy of ice thickness and density characterization.

[0134] In another preferred embodiment, such as Figure 3 and Figure 4 As shown, the cross-modal attention module uses the feature sequences of each spatial location point in the spatial geometric features as key vectors and value vectors; by calculating the multi-head self-attention similarity matrix between the query vector and the key vector, the attention weight coefficients of each feature element in the multimodal joint feature sequence to the spatial location point feature sequence are determined; the attention weight coefficients are used to perform weighted summation on the value vector to generate a three-dimensional voxel matrix of ice thickness and density aligned with the three-dimensional spatial coordinates.

[0135] Deploying a voxel morphological filter: The voxel morphological filter performs erosion and dilation operations sequentially on the three-dimensional voxel matrix to remove discrete, isolated noise voxels attached to the surface of the three-dimensional voxel matrix, specifically:

[0136] The voxel morphology filter uses a 3D structuring element to perform an opening operation on a 3D voxel matrix, i.e., first performing an erosion operation and then a dilation operation. The erosion operation involves sliding the 3D structuring element in the 3D voxel matrix, retaining the central voxel only when all voxels covered by the structuring element are effective icing voxels, otherwise discarding the central voxel, thereby removing small protrusions and isolated noise voxels from the surface of the 3D voxel matrix. The dilation operation involves sliding the 3D structuring element in the eroded voxel matrix, setting the central voxel as an effective icing voxel when there are effective icing voxels in the area covered by the structuring element, thereby restoring the shape of the main icing voxel cluster shrunk by the erosion operation, while retaining the effect of removing isolated noise voxels.

[0137] For each grid cell, the number of connected components in the voxel cluster after morphological filtering is counted. Specifically: a 3D connected component analysis algorithm is used to detect connected components in the voxel cluster corresponding to each grid cell. Based on the 6-neighbor or 26-neighbor connectivity rules, the connectivity between voxels is determined, and interconnected voxels are divided into the same connected component. The number of connected components in each grid cell and the voxel volume corresponding to each connected component are counted.

[0138] 6. Neighborhood connectivity rule: Two voxels are considered connected if they are adjacent in the three coordinate axes.

[0139] 26 Neighborhood Connectivity Rule: Two voxels are considered connected if they share any common vertex in 3D space;

[0140] When the number of connected components is greater than one, the largest main connected component voxel cluster is retained, and the suspended non-connected voxel clusters are removed. Specifically: when there are multiple connected components in a grid cell, the largest connected component is the main ice-covered voxel cluster attached to the jumper surface, and the other smaller non-connected components are suspended noise voxel clusters caused by laser point cloud density fluctuations. All voxels corresponding to the non-connected components are removed, and only the voxel clusters corresponding to the main connected components are retained; the final output is a morphologically filtered 3D voxel matrix.

[0141] The digital twin mesh for jumper catenaries is constructed by discretizing the actual catenary equations of the tension tower jumpers, dividing the jumpers into multiple mesh units with defined spatial topological connections. The specific steps are as follows:

[0142] Catenary morphology calculation: Based on the aforementioned catenary equation, combined with the actual span of the tension tower jumper, the coordinates of the suspension points at both ends, the jumper type and material parameters, the initial catenary morphology of the jumper is calculated.

[0143] Finite element discretization: The continuous catenary structure is divided into multiple one-dimensional beam elements using the finite element discretization method; each beam element is a mesh element, and adjacent mesh elements are connected through common nodes. All mesh elements and nodes together form a jumper catenary digital twin mesh with a defined spatial topological connection relationship.

[0144] Element attribute configuration: Each mesh element corresponds to a unique spatial location range and has defined material property parameters, including element length, cross-sectional area, elastic modulus and Poisson's ratio; each node contains three translational degrees of freedom, which are used to characterize the displacement changes of the node in three-dimensional space.

[0145] The filtered 3D voxel matrix and the jumper catenary digital twin mesh are obtained. The twin load mapping module executes the spatial bounding box collision detection algorithm to cluster and count the voxels falling within the spatial boundary of each mesh cell. The specific process is as follows:

[0146] Constructing spatial bounding boxes: The twin load mapping module constructs an axially aligned spatial bounding box for each mesh cell; the boundary of the spatial bounding box is determined by the extreme values ​​of the coordinates of the two endpoints of the mesh cell, that is, the x-axis range of the bounding box is the minimum and maximum x-coordinates of the two endpoints, the y-axis range is the minimum and maximum y-coordinates of the two endpoints, and the z-axis range is the minimum and maximum z-coordinates of the two endpoints, ensuring that the mesh cell is completely contained within the corresponding spatial bounding box;

[0147] Voxel attribution determination: For each voxel in the 3D voxel matrix, determine whether its center coordinates fall within the spatial bounding box of a certain mesh cell; if not, directly exclude the association between the voxel and the current mesh cell; if it falls within, further determine whether the voxel intersects with the spatial boundary of the mesh cell using the ray intersection method.

[0148] Voxel clustering: If a voxel intersects with or is completely within the spatial boundary of a mesh cell, it is classified into the corresponding mesh cell. For each mesh cell, after collision detection of all voxels, the number and total volume of all voxels classified into that mesh cell are counted, and the formula for calculating the total volume of voxels is as follows:

[0149] ,

[0150] in, For the first The total volume of ice-covered voxels corresponding to each grid cell To fall into the first The number of voxels within the spatial boundary of each grid cell. For the first The unit volume of an individual element.

[0151] After obtaining the total volume of icing voxels corresponding to each grid cell based on the above formula, the total volume of voxels is multiplied by the preset icing density parameter to obtain the local icing mass corresponding to that grid cell; wherein, the preset icing density parameter is set based on the typical type of icing of transmission lines and field measurement data.

[0152] Based on the gravitational acceleration vector and the normal vector of the mesh element, the local icing mass is transformed into a non-uniformly distributed load vector along the gravitational direction and perpendicular to the tangent direction of the mesh element, and the non-uniformly distributed load vector is applied to the nodes of the corresponding mesh element; the transformation process of the non-uniformly distributed load vector is achieved by the following formula:

[0153] ,

[0154] in, For the first Non-uniformly distributed load vector of each grid cell For the first Local icing quality of each grid cell It is the acceleration due to gravity. Let be the unit vector representing the vertically downward direction of gravity. It is the unit normal vector perpendicular to the tangent direction of the mesh cell.

[0155] Specifically, the local icing mass is decomposed into a component along the direction of gravity and a component perpendicular to the tangent of the grid cell. The vector sum of the two components is the non-uniformly distributed load vector of the grid cell. The twin load mapping module distributes the non-uniformly distributed load vector evenly to the two endpoint nodes of the grid cell according to the principle of equivalent nodal forces, thus completing the application of the non-uniformly distributed load vector on the jumper catenary digital twin grid.

[0156] In this embodiment, the volume of the voxel cluster of the main connected domain is used to replace the original voxel volume obtained by clustering in the calculation of local icing quality, so as to obtain the corrected local icing quality. Then, the non-uniformly distributed load vector is calculated based on the corrected local icing quality to eliminate the abnormal fluctuations in icing volume calculation caused by the fluctuation of laser point cloud density, thereby ensuring that the load calculation result is not affected by the noise voxels caused by the fluctuation of laser point cloud density. Table 3 shows the operation parameter configuration of the voxel morphology filter. This configuration is optimized for noise stripping and main icing morphology preservation of the three-dimensional voxel matrix to ensure that the filtering operation will not have a significant impact on the volume calculation of the main icing voxel cluster.

[0157] Table 3

[0158]

[0159] In Table 3, both the 3D erosion and dilation operations use 3×3×3 cubic structuring elements to ensure that the filtering operation has the same processing effect in all directions of 3D space. The number of iterations is set to 1 to avoid excessive shrinkage or expansion of the volume of the main ice-covered voxel cluster due to multiple operations. The connected component detection operation uses a 26-neighborhood connectivity rule, which can accurately identify all interconnected voxels in 3D space and ensure effective differentiation between the main connected component and the non-connected noise voxel cluster.

[0160] This embodiment constructs a digital twin mesh for jumpers based on the catenary equation, achieves precise matching between icing voxels and mesh cells through a spatial bounding box collision detection algorithm, converts voxel volumes into non-uniformly distributed load vectors, and eliminates noise voxels through voxel morphological filters and connected component analysis, thereby eliminating the influence of laser point cloud density fluctuations on the calculation of local icing quality and ensuring the accuracy of non-uniformly distributed load vector calculation.

[0161] In another preferred embodiment, such as Figure 5 and Figure 6 As shown, the nonlinear iterative solver uses the Newton-Raphson method for solving the problem. In each iteration step, the tangent stiffness matrix of the jumper catenary digital twin mesh is constructed based on the node coordinates of the current iteration step. The specific process is as follows:

[0162] Based on the grid node coordinates of the current iteration step, the strain tensor and deformation gradient of each grid element are calculated. Based on the material constitutive relation of the grid element, the element stiffness matrix of each grid element is calculated. Then, according to the topological connection relationship of the grid, all element stiffness matrices are assembled into the overall tangential stiffness matrix of the entire jumper catenary digital twin grid. The construction of the element stiffness matrix is ​​based on the mechanical properties of the beam element, considering the axial deformation, bending deformation and torsional deformation of the element, to ensure that the element stiffness matrix can accurately characterize the mechanical response characteristics of the element.

[0163] The nodal unbalanced force vector after applying a non-uniformly distributed load vector is multiplied by the inverse of the tangent stiffness matrix to obtain the nodal displacement increment. The calculation process of the nodal displacement increment is achieved by the following formula:

[0164] ,

[0165] in, For the first The node displacement increment vector of the iteration step. For the first The global tangent stiffness matrix of the iteration step. For the first The unbalanced force vector at each iteration step. Invert a matrix.

[0166] The node displacement increments are superimposed onto the node coordinates of the current iteration step, and the strain and stress tensors of each mesh element are updated according to the topological connectivity of the jumper catenary digital twin mesh. The update process of the strain and stress tensors is implemented by the following formula:

[0167] ,

[0168] ,

[0169] in, For the first Iteration step number The strain tensor of each mesh element For the first Iteration step number The geometric matrix of each grid cell. For the first Iteration step number The nodal displacement increment vector of each grid cell. For the first Iteration step number The stress tensor of each mesh element is the elastic constitutive matrix of the mesh element material.

[0170] When the norm of the nodal unbalanced force vector is less than a preset convergence threshold, the iteration terminates and the nodal coordinates and element stresses in the final stable state are output as the jumper morphology evolution result. Specifically, after updating the nodal coordinates and stress tensors in each iteration step, the nodal unbalanced force vector is recalculated. The nodal unbalanced force vector is the difference between the nodal external force vector and the nodal internal force vector. The L2 norm of the nodal unbalanced force vector is calculated. When the L2 norm is less than a preset convergence threshold, the iteration is considered converged, the iteration process is terminated, and the converged mesh nodal coordinates are output as the jumper morphology evolution result. At the same time, the stress tensor data of each mesh element are also output.

[0171] Meanwhile, the nonlinear iterative solver monitors the eigenvalues ​​of the deformation gradient matrix of each mesh element in real time during each iteration step. The calculation of the deformation gradient matrix and the distortion determination process are implemented through the following formula:

[0172] ,

[0173] ,

[0174] ,

[0175] in, For the first The deformation gradient matrix of each mesh element. It is the identity matrix. For the first The nodal displacement vector of each mesh element. Let be the initial coordinate vector of the grid cell. , , These are the three eigenvalues ​​of the deformed gradient matrix. The largest eigenvalue, It is the smallest eigenvalue. The distortion determination coefficient for the mesh element;

[0176] In each iteration step, the deformation gradient matrix of each grid cell is calculated, and the deformation gradient matrix is ​​decomposed into three eigenvalues. The ratio of the largest eigenvalue to the smallest eigenvalue is calculated to obtain the distortion determination coefficient. When the distortion determination coefficient exceeds the preset distortion threshold, the grid cell is determined to have severe grid distortion, and continuing the iteration may lead to iteration divergence.

[0177] For mesh cells that have mesh distortion, a local mesh reconstruction strategy is triggered: For a distorted mesh cell, one or more new nodes are inserted between the two endpoint nodes of the mesh cell to divide the original distorted mesh cell into multiple sub-mesh cells of uniform length. The number of sub-mesh cells is determined based on the degree of distortion of the original mesh cell. The higher the degree of distortion, the more sub-mesh cells are divided.

[0178] Based on the proportion of the volume of the sub-grid element to the volume of the original distorted grid element, the non-uniformly distributed load vector of the original grid element is distributed to each sub-grid element. At the same time, based on the coordinates of the adjacent nodes of the original distorted grid element, the coordinates of the newly inserted nodes are smoothed using the least squares method to ensure the continuity of the grid node coordinates and avoid iteration non-convergence caused by abrupt changes in local coordinates. After the local grid reconstruction is completed, the Newton-Raphson iteration process continues.

[0179] When calculating the skew angle data of the suspension string, the morphological stress solution module extracts the element stress at the end nodes connected to the suspension string fittings in the digital twin mesh of the jumper catenary. Specifically:

[0180] The two end nodes connected to the suspension string fittings in the digital twin mesh of the jumper catenary are located. The stress tensor data of the mesh elements to which the two end nodes belong are extracted. The element resultant force at the end node is calculated based on the stress tensor. The element resultant force is the integral result of the stress tensor on the cross section of the element.

[0181] The element stress at the end node is decomposed into a tension component along the jumper tangent and a lateral force component perpendicular to the jumper tangent. Specifically, based on the tangent and normal unit vectors of the mesh element to which the end node belongs, the element resultant force vector is projected onto the tangent and normal directions to obtain the tension component along the jumper tangent and the lateral force component perpendicular to the jumper tangent, respectively. The tangent unit vector is calculated based on the coordinate difference between the two endpoints of the mesh element, and the normal unit vector is calculated based on the cross product of the tangent unit vector and the gravity direction vector.

[0182] The spatial resultant tension vector of the suspension string is calculated based on the vector sum of its own gravity vector, tension component, and lateral force component. Then, the deflection angle of the suspension string is calculated based on the spatial resultant tension vector. The calculation process is achieved through the following formula:

[0183] ,

[0184] ,

[0185] in, The spatial resultant tension vector of the suspension string. Let be the vector of the suspended string's own gravity. Let be the tension component vector at the end node of the jumper. This represents the lateral force component vector at the jumper end node. Let be the spatial angle between the resultant tension vector in space and the vertically downward direction of gravity. Let be the unit vector representing the vertically downward direction of gravity. For the magnitude operation of vectors, This is the dot product operation for vectors.

[0186] By calculating the spatial angle between the spatial resultant tension vector and the vertically downward direction of gravity, and projecting this spatial angle onto the cross-section and longitudinal section of the transmission line, suspension string deflection angle data is generated, including the deflection angle along the line direction and the deflection angle across the line direction. Specifically, the spatial resultant tension vector is projected onto the longitudinal and cross-sectional planes of the transmission line, respectively. The longitudinal plane is the plane formed by the line direction and the vertical direction, and the cross-sectional plane is the plane formed by the line direction and the vertical direction. The angle between the projected resultant tension vector and the vertical direction is calculated to obtain the deflection angle along the line direction and the deflection angle across the line direction. These two deflection angles together constitute the suspension string deflection angle data.

[0187] During the calculation of the suspension string deflection angle data, the morphological stress solution module simultaneously acquires the wind field streamline data of the micro-topographic region where the tension tower is located. Specifically, the wind field streamline data is calculated based on numerical weather prediction data and topographic parameters of the micro-topographic region, including the distribution information of parameters such as wind speed, wind direction, and turbulence intensity in three-dimensional space. The temporal and spatial resolutions of the wind field streamline data are set based on the required calculation accuracy.

[0188] Based on the wind field streamline data, the aerodynamic drag vector of each grid cell is calculated. This aerodynamic drag vector is then superimposed on the non-uniformly distributed load vector to form a comprehensive dynamic load vector. The calculation process for the aerodynamic drag vector is achieved through the following formula:

[0189] ,

[0190] in, For the first The aerodynamic drag vector experienced by each grid cell. air density, This is the aerodynamic drag coefficient. For the first The windward area of ​​each grid cell This represents the relative velocity vector between the wind field and the grid cells. It is a unit vector in the direction of relative velocity.

[0191] Specifically, based on wind field streamline data, the wind speed and direction at the location of each grid cell are determined, the relative velocity between the wind field and the grid cell is calculated, the aerodynamic drag vector of each grid cell is calculated based on the above formula, the aerodynamic drag vector is vector-superimposed with the non-uniformly distributed load vector of the grid cell to obtain the comprehensive dynamic load vector, and the comprehensive dynamic load vector is applied to the corresponding grid cell node.

[0192] The Newton-Raphson method iteration is re-executed using the comprehensive dynamic load vector to extract the spatial resultant tension vector variation trajectory under dynamic time series. Specifically, based on wind field streamline data at different times, the comprehensive dynamic load vector corresponding to different times is calculated. For the comprehensive dynamic load vector at each time, the Newton-Raphson iteration is performed to obtain the grid node coordinates and element stress at each time, and then the spatial resultant tension vector of the suspension string at each time is calculated. The spatial resultant tension vectors at all times together constitute the spatial resultant tension vector variation trajectory under dynamic time series.

[0193] Extreme points are extracted from the trajectory of the spatial resultant tension vector change, and a deflection angle envelope space including the maximum longitudinal deflection angle and the maximum transverse deflection angle is constructed. The deflection angle envelope space is used as a component of the deflection angle data of the suspension string to characterize the maximum range of motion of the suspension string under wind-ice coupling conditions. The construction process of the deflection angle envelope space is achieved through the following formula:

[0194] ,

[0195] ,

[0196] ,

[0197] In the formula, The maximum deviation angle along the line, For the first The skew angle of the line at any given time, where , This represents the total duration of the time series. The maximum horizontal line skew angle, For the first The skew angle of the horizontal line at any given moment. It is the envelope space of the skew angle.

[0198] Specifically, the skewing angles along and across the line at each moment are extracted from the trajectory of the change of the spatial resultant tension vector. The maximum values ​​of the skewing angles along and across the line are calculated respectively. Based on the two maximum values, a skewing angle envelope space is constructed. This envelope space covers all possible values ​​of the skewing angle of the suspension string under wind-ice coupling conditions, which can accurately characterize the maximum range of motion of the suspension string and provide data support for the safety verification of transmission lines.

[0199] In this embodiment, the convergence control parameters of the nonlinear iterative solver are configured as shown in Table 4. This configuration is used to balance the computational efficiency and result accuracy of the iterative solution, while avoiding iterative divergence caused by large local deformations.

[0200] Table 4

[0201]

[0202] In Table 4, the convergence threshold for the unbalanced force norm is set as follows: To ensure sufficient accuracy of the iterative solution, the maximum number of iterations is set to 50 to avoid infinite loops in the iteration process. The distortion threshold is set to 5.0. When the ratio of the maximum eigenvalue to the minimum eigenvalue of a mesh cell exceeds 5.0, the mesh is considered to be severely distorted, triggering local mesh reconstruction. The maximum number of subdivided sub-cells is set to 4 to avoid excessive subdivision leading to a significant increase in computational load.

[0203] This embodiment employs the Newton-Raphson method for nonlinear iterative solution, accurately calculating the morphological evolution and stress distribution of the jumper under icing conditions. Based on the stress decomposition of the end nodes, the skew angle data of the suspension string is calculated. By real-time monitoring of mesh distortion and triggering local mesh reconstruction, iterative divergence caused by large local deformation is avoided. Combined with wind field streamline data, the aerodynamic drag vector is calculated to construct the skew angle envelope space under wind-ice coupling conditions, accurately characterizing the maximum range of motion of the suspension string.

[0204] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Any modifications, equivalent substitutions, or improvements made by those skilled in the art within the scope of the technology disclosed in this invention, based on the technical solution and inventive concept of the present invention, should be covered within the protection scope of this invention. Therefore, the protection scope of this invention should be determined by the scope of the claims.

Claims

1. A comprehensive analysis system for tension tower jumpers in complex scenarios, characterized in that, include: The multi-source data receiving module is used to receive visible light images, infrared thermal images, and laser point cloud data of tension tower jumpers collected by UAVs. A multimodal feature extraction module is used to extract surface texture features, temperature gradient features, and spatial geometric features of the jumper wire from the visible light image, the infrared thermal image, and the laser point cloud data, respectively, using a multimodal feature extraction network. The multimodal feature extraction network includes a visible light convolution branch, an infrared convolution branch, and a point cloud image convolution branch. The visible light convolution branch performs multi-scale convolution pooling operations on the visible light image, outputting a texture feature map that fuses local texture and global contour. The infrared convolution branch performs thermal radiation gradient difference calculations on the infrared thermal image, extracting the temperature difference boundary features between the iced area and the bare wire area as the temperature gradient feature. The point cloud image convolution branch constructs a topological adjacency graph of the laser point cloud data, aggregating the three-dimensional coordinate vectors and reflection intensity vectors of neighboring points through graph convolution kernels, outputting spatial geometric features including the three-dimensional spatial skeleton information of the jumper wire. The texture feature map, the temperature gradient features, and the spatial geometric features are projected onto a feature space of a unified dimension. The ice-covered voxel construction module is used to generate a three-dimensional voxel matrix representing the ice-covered distribution based on the extracted jumper surface texture features, temperature gradient features, and spatial geometric features through a cross-modal cross attention mechanism. The twin load mapping module is used to map the three-dimensional voxel matrix to a pre-established jumper catenary digital twin mesh, and calculate the non-uniformly distributed load vector according to the ice-covered voxel volume corresponding to each grid cell in the jumper catenary digital twin mesh. The morphological stress solution module is used to input the non-uniformly distributed load vector into a nonlinear iterative solver using the Newton-Raphson method, update the node coordinates and element stresses of the jumper catenary digital twin mesh, and output the jumper morphological evolution results and suspension string deflection angle data under icing conditions. The ice-covered voxel construction module implements the cross-modal cross-attention mechanism through the cross-modal cross-attention module in the following manner: The spatial location point feature sequence in the spatial geometric features is used as the query vector, and the multimodal joint feature sequence after concatenating the texture feature map and the temperature gradient feature is used as the key vector and value vector. By calculating the multi-head self-attention similarity matrix between the query vector and the key vector, the attention weight coefficient of each spatial location point feature sequence in the spatial geometric features to each feature element in the multimodal joint feature sequence is determined. The value vector is weighted and summed using the attention weight coefficient to generate the three-dimensional voxel matrix aligned with the three-dimensional spatial coordinates and representing the ice thickness and density. Alternatively, the cross-modal cross-attention mechanism can be implemented in the following manner: The cross-modal attention module uses the spatial location point feature sequence in the spatial geometric features as the key vector and value vector, and the multimodal joint feature sequence after concatenating the texture feature map and the temperature gradient feature as the query vector; by calculating the multi-head self-attention similarity matrix between the query vector and the key vector, the attention weight coefficient of each feature element in the multimodal joint feature sequence to the spatial location point feature sequence is determined; the value vector is weighted and summed using the attention weight coefficient to filter out the interference of texture features and spatial position misalignment caused by the drone shooting perspective shift, and the three-dimensional voxel matrix aligned with the three-dimensional spatial coordinates and representing the ice thickness and density is generated.

2. The comprehensive analysis system for tension tower jumpers in complex scenarios according to claim 1, characterized in that, The jumper catenary digital twin mesh is constructed by discretizing the actual catenary equation of the tension tower jumper, dividing the jumper into multiple mesh units with a defined spatial topological connection relationship; The twin load mapping module uses a spatial bounding box collision detection algorithm to cluster and statistically analyze the voxels in the three-dimensional voxel matrix that fall within the spatial boundaries of each grid cell. For each grid cell, the local icing mass of the corresponding grid cell is calculated by multiplying the voxel volume obtained from clustering with the preset icing density parameter. Based on the gravitational acceleration vector and the normal vector of the grid cell, the local icing mass is converted into a non-uniformly distributed load vector along the gravity direction and perpendicular to the tangent direction of the grid cell, and the non-uniformly distributed load vector is applied to the nodes of the corresponding grid cell.

3. The comprehensive analysis system for tension tower jumpers in complex scenarios according to claim 2, characterized in that, The nonlinear iterative solver uses the Newton-Raphson method to solve the problem; In each iteration step, the tangent stiffness matrix of the jumper catenary digital twin mesh is constructed based on the node coordinates of the current iteration step; The nodal unbalanced force vector after applying the non-uniformly distributed load vector is multiplied by the inverse of the tangent stiffness matrix to obtain the nodal displacement increment. The node displacement increment is superimposed on the node coordinates of the current iteration step, and the strain tensor and stress tensor of each mesh element are updated according to the topological connection relationship of the jumper catenary digital twin mesh. When the norm of the unbalanced force vector at the node is less than a preset convergence threshold, the iteration is terminated and the node coordinates and element stress in the final stable state are output as the result of the jumper morphology evolution.

4. The comprehensive analysis system for tension tower jumpers in complex scenarios according to claim 1, characterized in that, When calculating the skew angle data of the suspension string, the morphological stress solving module extracts the element stress at the end node connected to the suspension string hardware in the digital twin mesh of the jumper catenary. The unit stress at the end node is decomposed into a tension component along the jumper tangent and a lateral force component perpendicular to the jumper tangent. The spatial resultant tension vector of the suspension string is calculated based on the vector sum of its own gravity vector, the tension component, and the lateral force component. Calculate the spatial angle between the spatial resultant tension vector and the vertically downward gravity direction, and project the spatial angle onto the cross section and longitudinal section of the transmission line to generate the suspension string deflection angle data, which includes the deflection angle along the line direction and the deflection angle across the line direction.

5. The comprehensive analysis system for tension tower jumpers in complex scenarios according to claim 1, characterized in that, Before calculating the multi-head self-attention similarity matrix, the cross-modal cross-attention module introduces a modality credibility evaluation mechanism. For the texture feature components in the multimodal joint feature sequence, calculate the texture complexity entropy value of the corresponding pixel region. When the texture complexity entropy value is lower than the preset smoothing threshold, it is determined that the visible light features are degraded due to the light transmittance of the ice layer, and the proportion of the texture feature components in the query vector is reduced. For the temperature gradient feature components, the thermal radiation gradient variance is calculated. When the thermal radiation gradient variance is lower than a preset uniformity threshold, it is determined that the infrared features have degraded due to the homogenization of ambient temperature, and the proportion of the temperature gradient feature components in the query vector is reduced. The subsequent weighted summation calculation is performed based on the adjusted query vector.

6. The comprehensive analysis system for tension tower jumpers in complex scenarios according to claim 2, characterized in that, The twin payload mapping module deploys a voxel morphology filter before executing the spatial bounding box collision detection algorithm; The voxel morphology filter performs erosion and dilation operations on the three-dimensional voxel matrix in sequence to remove discrete isolated noise voxels attached to the surface of the three-dimensional voxel matrix. When calculating the local icing quality, for each grid cell, the number of connected components in the voxel cluster after morphological filtering is counted. When the number of connected components is greater than one, retain the largest main connected component voxel cluster and remove the dangling non-connected voxel clusters. The volume of the voxel cluster in the main connected domain is used to replace the original voxel volume obtained by clustering, and is then used in the calculation of the local icing mass.

7. The comprehensive analysis system for tension tower jumpers in complex scenarios according to claim 3, characterized in that, In each iteration step, the nonlinear iterative solver monitors the eigenvalues ​​of the deformation gradient matrix of each mesh cell in real time. When the ratio of the maximum eigenvalue to the minimum eigenvalue of the deformation gradient matrix of any mesh cell exceeds a preset distortion threshold, the mesh cell is determined to have mesh distortion. For the mesh cell that has undergone mesh distortion, a local mesh reconstruction strategy is triggered. Based on the adjacent nodes of the mesh cell, new nodes are inserted and the distorted mesh cell is divided into multiple sub-mesh cells. The non-uniformly distributed load vector of the original distorted mesh element is distributed to each of the sub-mesh elements according to the principle of volume equivalence, and the node coordinates of the jumper catenary digital twin mesh are updated smoothly using the least squares method.

8. The comprehensive analysis system for tension tower jumpers in complex scenarios according to claim 4, characterized in that, During the calculation of the suspension string deflection angle data, the morphological stress solving module simultaneously acquires the wind field streamline data of the micro-topography area where the tension tower is located. Based on the wind field streamline data, the aerodynamic drag vector of each grid cell is calculated, and the aerodynamic drag vector is superimposed on the non-uniformly distributed load vector to form a comprehensive dynamic load vector. The Newton-Raphson method iteration is re-executed using the comprehensive dynamic load vector to extract the spatial resultant tension vector change trajectory under dynamic time sequence. The extreme points in the trajectory of the change of the spatial resultant tension vector are extracted, and a deflection angle envelope space containing the maximum longitudinal deflection angle and the maximum transverse deflection angle is constructed. The deflection angle envelope space is used as a component of the output of the deflection angle data of the suspension string to characterize the maximum range of motion of the suspension string under wind-ice coupling conditions.

Images