A transmission tower tilt model detail optimization method

Abstract

The present application relates to a power transmission tower tilt model detail optimization method, belongs to the field of three-dimensional reconstruction of power transmission tower, solves the problem of insufficient accuracy of power transmission tower tilt model reconstruction based on oblique photography image in the prior art. The specific steps include: generating a power transmission tower original point cloud from multi-view power transmission tower images, and extracting a grid-based tower effective point cloud therefrom, and extracting a geometric feature sequence from the tower effective point cloud; inputting the geometric feature sequence into a pre-trained classification network to obtain the deformation probability of the local components of the tower body; Gaussian modeling is performed on the tower effective point cloud to obtain a plurality of Gaussian cells and corresponding Gaussian parameters, the Gaussian overlap degree of adjacent Gaussian cells is calculated based on the Gaussian parameters, and the Gaussian parameters are adaptively updated based on the Gaussian overlap degree and the deformation probability of the local components to obtain an optimized tower center line point set; based on the optimized tower center line point set, entity modeling is performed to obtain a detail-optimized power transmission tower tilt model, and the tilt modeling accuracy is improved.

Description

Technical Field

[0001] This invention relates to the field of three-dimensional reconstruction technology of transmission towers, and in particular to a method for optimizing the details of a tilt model of a transmission tower. Background Technology

[0002] Oblique photogrammetry modeling is a cutting-edge technology that has emerged in the international surveying and mapping field in recent years. Through multi-angle image acquisition, 3D point cloud reconstruction, and texture mapping, it can quickly generate highly realistic and accurate 3D models of real-world scenes. This technology has been widely applied in fields such as digital city construction, engineering surveying, and cultural heritage protection, and compared to traditional modeling methods, it features high efficiency, high accuracy, and low cost.

[0003] However, oblique photogrammetry modeling technology still has significant limitations in its ability to reproduce details when dealing with objects with complex geometric features, such as power transmission towers. Power transmission towers and lines, as critical infrastructure for power transmission, are assembled from numerous small metal components such as angle irons, bolts, and gaskets. Detailed modeling of these complex components is crucial for project management, quality supervision, and engineering analysis. However, in most actual modeling processes, oblique photogrammetry struggles to accurately capture the complete geometric information of these small components, leading to distortion or even patchy images, failing to accurately reproduce the clear impact features of these delicate components. Furthermore, the unique geometry of power transmission towers, especially after the completion of the overhead line construction phase, means that when drones carrying multiple cameras take photos from different angles, the shooting angles and trajectories are affected, resulting in a decrease in both the quantity and quality of the photos, inevitably leading to a corresponding decline in the quality of the 3D reconstruction. Summary of the Invention

[0004] Based on the above analysis, the present invention aims to provide a method for optimizing the details of a transmission tower tilt model, in order to solve the problem of insufficient accuracy in the reconstruction of a transmission tower tilt model based on oblique photographic images in the prior art.

[0005] The objective of this invention is mainly achieved through the following technical solutions: This invention provides a method for optimizing the details of a transmission tower tilt model, comprising the following steps: The original point cloud of the transmission tower is generated from multi-view images of the transmission tower, and a gridded effective point cloud of the tower body is extracted from it. Geometric feature sequences are extracted from the effective point cloud of the tower body according to the spatial position of the grid. The geometric feature sequence is input into a pre-trained classification network to obtain the deformation probability of local components of the tower. Gaussianization modeling is performed on the effective point cloud of the tower to obtain several Gaussian cells and corresponding Gaussian parameters. Based on the Gaussian parameters, the Gaussian overlap of adjacent Gaussian cells is calculated in different regions. Based on the Gaussian overlap and the deformation probability of local components, the Gaussian parameters are adaptively updated to obtain the optimized centerline point set of the tower. Based on the optimized tower centerline point set, solid modeling is performed to obtain a detailed optimized transmission tower tilt model.

[0006] Furthermore, the Gaussian parameters include the center point coordinates and the covariance matrix; adaptively updating the Gaussian parameters includes: The covariance matrix of each Gaussian cell is adaptively updated based on the Gaussian overlap and the deformation probability of local components. Based on the position of the center point of each Gaussian cell to be updated in the tower structure, the differentiated neighborhood fusion range corresponding to each Gaussian cell to be updated is determined; based on the distances between each other Gaussian cell within the neighborhood fusion range and the corresponding Gaussian cell to be updated, combined with the deformation probability of local components, the dynamic weights corresponding to each other Gaussian cell are determined; based on the center point coordinates of each other Gaussian cell and the corresponding dynamic weights, spatial weighting is performed to obtain the updated center point coordinates corresponding to each Gaussian cell to be updated.

[0007] Furthermore, the dynamic weights decrease exponentially with respect to the square of the distances to other Gaussian cells within the neighborhood fusion range; the dynamic weights of each Gaussian cell are determined based on the following constraints: Based on the deformation probability of local components, the confidence level of the center point corresponding to the coordinates of the center point of each Gaussian cell is determined; if the Gaussian cell to be updated is located in the edge region, the neighborhood fusion range is expanded, and the dynamic weights of other Gaussian cells within the neighborhood fusion range are proportional to the confidence level of the corresponding center point. If the Gaussian cell to be updated is located in a smooth region, then the neighborhood fusion range is shrunk.

[0008] Furthermore, the covariance matrix corresponding to each Gaussian cell is adaptively updated, including: Calculate the local curvature of the region where each Gaussian cell is located; The update weights corresponding to the covariance matrix of each Gaussian cell are preset. The update weights are directly proportional to the Gaussian overlap of each Gaussian cell and inversely proportional to the corresponding local curvature. The covariance matrix of each Gaussian cell is expanded or reduced based on the corresponding update weights.

[0009] Furthermore, the optimized set of tower centerline points is obtained based on the following process: Based on the Gaussian parameters, the critical contact area and non-critical contact area of ​​the tower structure are identified, and the Gaussian overlap of the corresponding areas is calculated. An overlap loss is constructed based on the Gaussian overlap degree. The optimized gradient of each center point is calculated based on the overlap loss. The coordinates of the center points of each Gaussian cell are iteratively and adaptively updated along the direction of each negative optimized gradient. In each iteration, the slope of the center line is calculated based on the updated center point coordinates. The covariance matrix is ​​adaptively updated by the rate of change of the center line slope. The total energy function is calculated based on the updated center point coordinates and the covariance matrix. The iteration process is completed by minimizing the total energy function to obtain the optimal Gaussian parameters of each Gaussian cell. The optimized tower center line point set is obtained from the center point coordinates of the optimal Gaussian parameters.

[0010] Furthermore, based on various Gaussian parameters, the critical and non-critical contact areas of the tower body are identified, including: Traverse all Gaussian cells and calculate the Euclidean distance between the center points of any two Gaussian cells to obtain the corresponding spatial distance. If the spatial distance is less than a preset distance threshold and the principal direction of the corresponding covariance matrix points to the other, then the two Gaussian cells are identified as critical contact areas; otherwise, the two Gaussian cells are identified as non-critical contact areas.

[0011] Furthermore, the Gaussian overlap of the corresponding regions is calculated, including: Based on the center point coordinates and covariance matrix of each Gaussian cell in the critical contact area, the weighted sum of the overlap volume of the Gaussian probability density function and the Bach coefficient is calculated to obtain the Gaussian overlap of the critical contact area. Based on the center point coordinates and covariance matrix of each Gaussian cell in the non-critical contact area, the weighted sum of the intersection volume of the envelope boxes and the spatial distance of all Gaussian cells is calculated to obtain the Gaussian overlap of the non-critical contact area.

[0012] Furthermore, the overlap loss is expressed as: , in, The overlap loss value is used to measure the degree of overlap between all Gaussian cells; Indicates the first Gaussian cells and the Gaussian cells Gaussian overlap; This indicates the preset target overlap threshold.

[0013] Furthermore, the total energy function is expressed as: , in, This represents the total energy value. For the regularity constraint term of the covariance matrix of Gaussian cells based on local curvature constraints; The deviation between the coordinates of the Gaussian cell center point and the center line point of the design blueprint; This represents the overlap error between adjacent Gaussian cells, i.e., the overlap loss. These are the corresponding energy term weight coefficients.

[0014] Furthermore, the classification network is an LSTM classification network; the deformation probability of local components of the tower body is obtained, including: The geometric feature sequence is decomposed into multiple IMF components using variational mode decomposition. Singular value decomposition is then performed on each IMF component, and a preset number of singular values ​​are extracted to form a feature vector, thus obtaining the singular value sequence corresponding to each IMF component. Each singular value sequence is input into an LSTM classification network, and the deformation probability of each local component of the tower is obtained from the output.

[0015] Compared with the prior art, the present invention can achieve at least one of the following beneficial effects: 1. This invention proposes a Gaussian probabilistic representation for transmission tower point clouds acquired from multi-view images. It calculates the Gaussian overlap of each Gaussian cell through differentiated partitioning to estimate deformation, and combines this with deformation probabilities identified by a classification network to adaptively fuse and update Gaussian parameters to obtain the optimized centerline point of the tower. By combining traditional point cloud technology with Gaussian splashing techniques and an adaptive update strategy to optimize Gaussian parameters, environmental interference is compensated for, noise resistance is improved, and the accuracy of the transmission tower tilt model is enhanced. This provides a lightweight and highly robust solution for safety monitoring of transmission towers in complex environments.

[0016] 2. The covariance matrix is ​​updated collaboratively using Gaussian overlap and local curvature, and an adaptive update mechanism for updating the Gaussian center position is based on the edge-smooth region differential neighborhood fusion strategy. This not only prevents the loss of contour features but also reduces the computational cost of smooth regions. At the same time, the update of edge Gaussian points is constrained by the high-confidence centerline to reduce noise interference and obtain a high-precision tower centerline point set, providing a reliable model foundation for tower deformation, tilt assessment, and detailed reconstruction of tower tilt model.

[0017] 3. By introducing an intelligent identification mechanism for critical / non-critical contact areas, the inherent defects of traditional point cloud deformation analysis, such as huge computational consumption and false alarms, are overcome. This mechanism achieves precise allocation of computing resources, improving the sensitivity of core node deformation detection while providing physically reliable prior knowledge for subsequent differentiated fusion optimization.

[0018] In this invention, the above-described technical solutions can be combined with each other to achieve more preferred combinations. Other features and advantages of this invention will be set forth in the following description, and some advantages may become apparent from the description or be learned by practicing the invention. The objects and other advantages of this invention can be realized and obtained from what is particularly pointed out in the description and drawings. Attached Figure Description

[0019] The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Throughout the drawings, the same reference numerals denote the same parts.

[0020] Figure 1 This is a flowchart of the detailed optimization method for the tilt model of the transmission tower according to an embodiment of the present invention; Figure 2 The attached diagram shows the Gaussian overlap evaluation process according to an embodiment of the present invention. Figure 3 The accompanying drawing shows the Gaussian weighted fusion process according to an embodiment of the present invention; Figure 4 The attached figure shows the post-processing flow after iterative optimization in an embodiment of the present invention. Detailed Implementation

[0021] Preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings, which form part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, but are not intended to limit the scope of the present invention.

[0022] Example 1 A specific embodiment of the present invention discloses a method for detailed optimization of a transmission tower tilt model, such as... Figure 1 As shown, it includes the following steps: Step S1: Generate the original laser point cloud of the transmission tower from the multi-view transmission tower image, extract the gridded effective point cloud of the tower body from it, and extract the geometric feature sequence of the effective point cloud of the tower body according to the grid spatial position. Step S2: Input the geometric feature sequence into a pre-trained classification network to obtain the deformation probability of local components of the tower. Step S3: Gaussianize the effective point cloud of the tower body to obtain several Gaussian cells and corresponding Gaussian parameters. Based on the Gaussian parameters, calculate the Gaussian overlap of adjacent Gaussian cells in different regions. Based on the Gaussian overlap and the deformation probability of local components, adaptively update the Gaussian parameters to obtain the optimized centerline point set of the tower body. Step S4: Based on the optimized tower centerline point set, perform solid modeling to obtain a detailed optimized transmission tower tilt model.

[0023] Using the above method, with Gaussian cells as the unified data carrier, the deformation probability identified by the classification network, the Gaussian overlap calculated by differential partitioning, and the adaptive update of Gaussian parameters are organically integrated. Through unsupervised optimization of Gaussian parameters, the centerline point set constrained by global consistency is finally obtained. This improves the geometric consistency of tower components, enables detailed completion of the transmission tower model, and enhances the accuracy of the reconstruction of the transmission tower tilt model.

[0024] Specifically, in step S1, the original laser point cloud of the transmission tower is obtained from multi-angle images of the transmission tower taken by the drone, and the original laser point cloud is processed into a grid to extract the effective point cloud of the transmission tower body. Geometric feature sequences are extracted from the effective point cloud of the tower body according to the grid spatial position. The specific steps are as follows: S11. Kalman filtering is applied to the original laser point cloud for denoising, and dynamic noise suppression is performed on the acquired noisy point cloud to obtain a denoised point cloud. By fusing the system dynamic model with the noisy observation data, the optimal estimate of the signal is achieved, expressed as: , , in: express Time-state estimation, point coordinates; B represents the state transition matrix as an identity matrix, and B represents the control input matrix. For control vectors; , This represents the covariance between process noise and observation noise; for Time-based process noise, This represents a normal distribution with a mean of 0 and a covariance of Q. for Observational noise at any given moment; This represents the measurement vector at time k, corresponding to the noisy point cloud data collected by the sensor.

[0025] Through iterative state prediction and observation correction, the system gradually approximates the real signal, eliminates environmental noise such as wind disturbance and sensor jitter, and preserves the details of the tower structure. S12. Segment the denoised point cloud using elevation histograms, divide the denoised point cloud into grids, count the elevation distribution peak of each grid, distinguish between transmission towers (continuous elevation) and vegetation / conductors (discrete elevation), extract the effective point cloud of the transmission tower body and the corresponding elevation value sequence, and provide clean input for subsequent modeling by separating the target structure from the interference. For example, firstly, the denoised point cloud is divided into grids along the main direction of the transmission line; then, elevation histograms are statistically analyzed for all denoised point clouds within each spatial grid, and local maxima are extracted from each elevation histogram to obtain the elevation sequence corresponding to each spatial grid. Finally, by setting a pre-defined elevation threshold, ground points and non-ground points are filtered in each elevation sequence, and the transmission tower elevation is extracted. Typically, the peak elevation of the transmission tower point cloud exhibits continuous, smooth, and orderly changes, while vegetation height is random and discontinuous, with dramatic fluctuations in peak elevation between adjacent grids. Although the conductor is spatially continuous, the number of point clouds within a single grid is far less than that of the tower, resulting in low point cloud density and a lack of continuous elevation distribution characteristics. Based on these characteristics, the target transmission tower point cloud is separated from the segmented non-ground points by analyzing the maximum vertical interval of the point set. The expression for separating target features from a complex environment is as follows: , in: Indicates the first Line number The grid cells of a column store point data that conforms to the specified spatial range; Indicates the horizontal position (precision), vertical position (dimension), and elevation value of a point within a grid cell; This represents the side length of the grid, that is, the fixed length of each grid cell in the X and Y directions; ) is the first Line number The coordinates of the grid cells in the column on the XY plane.

[0026] S13. Based on the images at different inspection times, extract the geometric feature sequence of the effective point cloud of the gridded tower body at the corresponding time, specifically including features such as elevation, surface curvature, and normal vector, to reflect the changes of each feature on the spatial grid at different times.

[0027] Specifically, in step S2, the extracted geometric feature sequence is input into a deep learning classification network to extract local deformation features of key contact areas, identify potential risk components, and obtain the probability of local structural deformation of the tower body, thus supplementing the micro-tilt risk not covered by the Gaussian model. The specific process is as follows: S21. The geometric feature sequence is decomposed into multiple intrinsic mode function components (IMFs) using variational mode decomposition-singular value decomposition (VMD-SVD). The decomposition parameters are adaptively selected by an optimization algorithm. Singular value decomposition is performed on each IMF component, and the first few singular values ​​are extracted to form a feature vector, resulting in the singular value sequence corresponding to each IMF component. This captures modal changes in tower vibration and characterizes stiffness anomalies. The VMD-SVD feature extraction process is as follows: , , in: Let m represent the m-th modal component (IMF), and let m represent the sub-signal after signal decomposition. This represents the center frequency of the m-th mode; It represents the partial derivative with respect to time, used to calculate the instantaneous rate of change of a signal; This represents the Dirac impulse function, used to construct the kernel function of the Hilbert transform; The imaginary unit represents the frequency domain; Convolution operations combine kernel functions with modal components; The frequency shift factor indicates that the modal spectrum is shifted to the baseband. This represents the mode matrix, with each column representing a mode component derived from VMD decomposition. Represents a left singular vector matrix; Represents a singular value diagonal matrix. Construct a function for a diagonal matrix; This represents the transpose of the right singular vector matrix, and an orthogonal basis for the time sample space of row vectors; This represents the first singular value, reflecting the weight (energy percentage) of the corresponding mode.

[0028] S22. The classification network is an LSTM classification network. Each singular value sequence is input into the LSTM classification network, and by capturing the temporal dependence, the micro deformation index is output. The higher the index, the more the geometric features of the region deviate from the normal tower structure, and the greater the risk of micro tilting.

[0029] It should be noted that before optimizing the details of the transmission tower tilt model, a trained LSTM classification network is obtained through offline training. Before training, the component categories of key contact areas such as tower foot foundations, crossarms, and bolts in different locations are manually labeled in the effective point cloud of the tower body. Random rotation, translation, and changes in brightness are used to enhance data diversity and generate supervised learning training samples to support the training of the component classification model. The LSTM classification network learns temporal dependencies through gating units, and the processing procedure is as follows: , , , , , , in: Indicates the output of the forget gate; The weight matrix of the forget gate has the shape (hidden_size, hidden_size + input_size); It represents the hidden state of the previous time step, stores historical information, and has a shape (hidden_size). This represents the current input data, and its shape (input_size). For the bias term of the forget gate; The Sigmoid activation function is used. Its function is to control which information in the cellular state needs to be forgotten; The output of the input gate, ranging from [0,1], determines the candidate state. The update ratio; This represents the weight matrix of the input gate; This represents the bias term of the input gate; Its function is to filter out valuable information from the current input and update the cell state; Indicates the output of the output gate; This represents the weight matrix of the output gate; This represents the bias term of the output gate; Its function is to determine the output of the hidden state at the current moment; Represents the candidate state, ranging from [-1, 1], and stores the new information of the current input; This is the weight matrix for the candidate states; Bias terms representing candidate states; Represents the hyperbolic tangent activation function; Its function is to generate states to be added to the cell. Potential new information within; This represents the hidden state at the current moment, serving as the network output and the input for the next moment; This indicates a non-linear transformation of the cell state, limiting the output range; Its function is to generate the final output at the current time step and pass it to the next time step.

[0030] Specifically, in step S3, the effective point cloud of the tower is Gaussianized and modeled to obtain several Gaussian cells and corresponding Gaussian parameters. The consistency of point cloud distribution among adjacent components is quantified by Gaussian overlap. By evaluating the degree of tower structural deformation, the Gaussian parameters are adaptively fused and optimized. This eliminates point cloud drift caused by structural deformation while accurately preserving the sharp edges and assembly features of the tower structure. The process of finally obtaining the optimized centerline point set of the tower is as follows: S31. Establish geometric constraints using the symmetry of the transmission tower, and calculate the ideal centerline of the tower body based on the transmission tower design blueprint data. The geometric constraints include planar constraints, vertical constraints, and isoangular constraints; among which... Planar constraints are applied by fitting plane equations to the four sides of the tower, ensuring that each side conforms to linear geometric relationships, defining the spatial boundaries of the tower body, and replacing the manual setting of projection planes; the planar constraints are expressed as: , in: , , They represent the first The X, Y, and Z components of the plane parameter normal vector of a plane, and the normal vector. Perpendicular to the plane in which it is located; , , These represent the coordinates of any point in three-dimensional space; This represents the constant term in the equation of the plane, which does not pass through the origin.

[0031] Vertical constraints force adjacent tower surface normals to be orthogonal, eliminating errors from non-perpendicular structures, preserving the prism properties of the tower, and reducing computational degrees of freedom. Vertical constraints are expressed as: , in, , These represent the X and Y components of the normal vectors of two adjacent planes, respectively. Anisotropic constraints are expressed as: , in, The denominator represents the preset fixed angle, determined according to the parameters in the transmission tower design blueprint; the denominator represents the magnitude of the normal vector XY.

[0032] Three sets of constraint equations are combined, and the tower profile is solved using the extracted point cloud elevation and normal vectors. Specifically, a baseline centerline is generated by connecting the diagonals of the tower top and bottom, and the slope of the centerline is calculated to obtain the tower tilt angle. By constructing a physical driving model based on geometric constraints, the calculation of the transmission tower tilt is simplified.

[0033] S32. Gaussian cell modeling is performed on the extracted effective point cloud of the tower body. Using the points on the ideal centerline of the tower body generated by geometric constraints as a reference, the effective 3D point cloud of the transmission tower body is converted into Gaussian ellipsoidal probability units. That is, the effective point cloud of the tower body is divided into several Gaussian cells, each corresponding to a set of Gaussian parameters, including the center point coordinates and the shape covariance matrix. When modeling based on the points on the ideal centerline, the center point coordinates of each Gaussian cell are initialized to the coordinates of the corresponding point on the ideal centerline. The Gaussian cell modeling is as follows: , in: Point In the The probability density value in a Gaussian cell; Represents the coordinates of any point in three-dimensional space; This represents the center point of a Gaussian cell; The covariance matrix describes the shape and orientation of the Gaussian cell; The exponential function maps the Mahalanobis distance to probability values. Mahalanobis distance, measuring point To the center Weighted distance.

[0034] S33. Analyze the adjacency relationship of Gaussian cells based on each Gaussian parameter, identify the key contact area and non-key contact area of ​​the tower body, calculate the Gaussian overlap of the corresponding area, quantify the consistency of point cloud distribution of adjacent components, and use it for subsequent evaluation of tower body structural deformation to adaptively adjust the Gaussian parameters. It should be noted that critical contact areas are spatial connection areas between adjacent components that should be physically and mechanically closely fitted and bear the task of force transmission. Examples include bolted connections between the main tower structure and diagonal members, hinged connections between crossarms and the tower structure, and anchorage areas of the tower foot plate and foundation bolts. These areas should theoretically have a high degree of overlap; any decrease in overlap indicates loose bolts or weld cracks, posing a significant structural safety hazard. Non-critical contact areas include two parallel auxiliary angle steels (although spatially close, they are anchored to different main members and do not directly transmit force) and cable supports inside the tower structure used for wiring (even slight misalignment does not affect the safety of the main structure). By introducing an intelligent identification mechanism for critical / non-critical contact areas, the reliability and computational efficiency of subsequent tower deformation assessments can be improved.

[0035] Specifically, such as Figure 2 As shown, based on the center point coordinates and shape covariance matrix of each Gaussian cell, the actual occupancy and interaction state of two adjacent Gaussian cells in space are analyzed to determine whether they are critical contact areas: by calculating the Euclidean distance between the center points of two Gaussian cells, the spatial distance is obtained. When the spatial distance is less than a preset distance threshold and the main direction of the shape covariance matrix points to the other, the two Gaussian cells are determined to be critical contact areas.

[0036] For critical contact areas, a strict overlap standard is adopted. The overlap volume and Bhattacharyya coefficient of the Gaussian probability density function are accurately calculated based on Gaussian parameters. After normalization and weighted summation, the Gaussian overlap degree of the critical contact area is obtained to capture subtle looseness or misalignment. Conversely, for non-critical contact areas, which have low overlap, a relaxed overlap standard is adopted. The Gaussian overlap degree of the non-critical contact area is obtained by quickly evaluating the overlap volume of the envelope boxes of all Gaussian cells and the center distance, thus saving computing power and achieving a coarse screening for collision prevention.

[0037] S34. The Gaussian overlap and the deformation probability of local components are introduced into the fusion process as prior weights. The adaptive update of the Gaussian parameters includes the update of the center point coordinates of each Gaussian cell and the update of the shape covariance matrix.

[0038] On one hand, the covariance matrices are dynamically scaled based on both Gaussian overlap and local curvature. This includes expanding the covariance in stable regions with high Gaussian overlap and low local curvature to smooth the surface; and reducing the covariance in deformable or edge regions with low Gaussian overlap or high curvature to lock angular features and adaptively protect the tower's edge structure. Based on these constraints, update weights are pre-set for each Gaussian cell's covariance matrix, making the update weights proportional to the Gaussian overlap of each cell and inversely proportional to the corresponding local curvature. Based on these update weights, each Gaussian cell is expanded or reduced accordingly. The dynamic adjustment of covariance is expressed as: , , in: This is the updated covariance matrix; Here is the covariance matrix to be updated; Gaussian cells The corresponding Gaussian overlap; Let be the curvature coefficient of Gaussian cell j. , These are the corresponding adjustment weights (usually set as follows). ), and strictly constrain when High and When the scaling factor is too low, it is greater than 1 to expand the covariance smoothing surface. Low or When the scaling factor is too high, it is less than 1 to reduce the covariance and protect the angular features; This represents the scaling factor, which needs to be tuned based on gradient feedback later. Represents Gaussian cells The local curvature at a given point is estimated by calculating the difference in normal vectors. Represents Gaussian cells The normal vector; These are Gaussian cells. The normal vector of the next adjacent point; its overall function is to avoid excessive smoothing that leads to blurred edges.

[0039] On the other hand, when updating the center point coordinates of any Gaussian cell, it is determined whether the center point of the Gaussian cell is located in the edge region or a smooth region of the tower. Based on its location, the differentiated neighborhood fusion range of each Gaussian cell to be updated is determined. Based on the distances between each other Gaussian cell within the neighborhood fusion range and the Gaussian cell to be updated, the dynamic weights of each other Gaussian cell are determined. A spatially weighted update is performed based on the center point coordinates of each other Gaussian cell and its corresponding dynamic weights to obtain the updated center point coordinates of each Gaussian cell to be updated. Here, "other Gaussian cells" refers to all Gaussian cells within the neighborhood fusion range except the Gaussian cell to be updated. The dynamic weights decay exponentially with the square of the distances to each other Gaussian cell within the neighborhood fusion range. Furthermore, the dynamic weights also satisfy the following constraints: based on the deformation probability of local components, the confidence level of the center point coordinates of each Gaussian cell relative to the ideal center line is determined; if the Gaussian cell to be updated is located in the edge region, the neighborhood fusion range is expanded, and the higher the confidence level of the center point of each Gaussian cell within the neighborhood fusion range, the greater the corresponding dynamic weight; if the Gaussian cell to be updated is located in the smooth region, the neighborhood fusion range is shrunk.

[0040] For example, the adaptive fusion process involves iterating through all Gaussian cells and performing the corresponding Gaussian parameters as follows: Figure 3 As shown, for any Gaussian cell center point to be updated, firstly, the set of Gaussian cell center points within a preset initial neighborhood fusion range centered on it is obtained, referred to as the neighborhood point set; then, by analyzing local geometric features, it is identified whether the center point is located at the edge of a structure or in a smooth region, thus determining the differential neighborhood fusion range. Within the differential neighborhood fusion range, dynamic weights corresponding to each neighboring Gaussian cell are constructed based on the Euclidean distance between each neighboring Gaussian cell center point and the Gaussian cell center point to be updated; where, the neighboring Gaussian cells are all Gaussian cells within the neighborhood fusion range except for the Gaussian cell to be updated. For each Gaussian cell, the closer the neighboring Gaussian cell is to the Gaussian cell to be updated, the higher its corresponding dynamic weight. For the center point coordinates of the Gaussian cell to be updated in the edge region, the neighborhood fusion range is expanded and the dynamic weight of the high-confidence center line point is increased to force the actual deviation at the edge and prevent feature loss. For the center point to be updated in the smooth region, i.e., the non-edge region, the neighborhood fusion range is shrunk to reduce the computational cost. Finally, the spatial weighted average is performed based on the center point coordinates of each Gaussian cell within the neighborhood fusion range and the corresponding dynamic weights to obtain the corresponding updated center point coordinates.

[0041] It is important to note that, based on the characteristic that stronger edge features result in greater local curvature, to expand the neighborhood fusion range of Gaussian cells located in edge regions, local curvature is introduced into dynamic weights. By slowing down the decay rate of weights with distance, the effect of expanding the neighborhood fusion range is achieved. Based on this, the updated center point coordinates are represented as: , in: Indicates the center point The new coordinates after fusion; Indicates the Gaussian cell center point to be updated. The set of neighborhood points; Represents the center point of the neighborhood Gaussian cell. The original coordinates; Indicates the center point The original coordinates; Point With point The square of the Euclidean distance between them; For the center point confidence, and , This represents the deformation probability of the corresponding local component. The center point of the neighboring Gaussian cell The projected distance to the ideal centerline; For geometric consistency tolerance coefficient, control The rate at which the value increases as the confidence level of the point increases.

[0042] S35. Construct the overlap loss from the Gaussian overlap, calculate the partial derivative of the overlap loss with respect to the center point of each Gaussian cell, and obtain the corresponding optimization gradient. Based on the optimization gradient, perform unsupervised optimization using the gradient descent method. That is, the optimization gradient indicates the adjustment direction and magnitude required for the center coordinates and shape covariance matrix of each Gaussian cell. Finally, the optimal Gaussian parameters of each Gaussian cell are obtained through optimization.

[0043] Specifically, such as Figure 4 As shown, the center point coordinates of each Gaussian cell and its associated covariance matrix are used as parameters to be optimized. An overlap loss is constructed using Gaussian overlap. The optimization gradient of each center point is calculated based on the overlap loss, and the center point coordinates of each Gaussian cell are iteratively and adaptively updated along the directions of the negative optimization gradients. In each iteration, the slope of the centerline is calculated based on the updated center point coordinates. The direction and scale of the principal axis of the covariance are dynamically determined by the rate of change of the centerline slope, and the covariance matrix is ​​adaptively updated, thereby achieving adaptation to local structural features. The total energy function is then refitted based on the updated center point coordinates and the covariance matrix, and the iteration process is completed by minimizing the total energy function. Specifically, the overlap error is obtained based on the Gaussian overlap of adjacent Gaussian cells, and an overlap loss is constructed based on the overlap error to evaluate the tower deformation. The overlap loss is expressed as: , in: The loss value is used to measure the degree of overlap between all Gaussian cells; Indicates the first The Gaussian cell and the first Gaussian overlap of a Gaussian cell; This indicates the preset target overlap threshold, which is dynamically set based on the region attributes. For example, the critical contact area takes a high positive value of the theoretical overlap volume, while the non-critical contact area takes 0 to distinguish them.

[0044] The fine-tuning of gradient descent during the iteration process is as follows: , in: Indicates the first During the nth iteration The center coordinates of a Gaussian cell; This represents the learning rate and controls the step size for updating parameters. Represents the overlapping loss function pair The gradient points in the direction of the fastest decrease in loss; This indicates the current iteration number; the method optimizes the position of Gaussian cells through gradient descent, minimizing the overlap between cells.

[0045] If the change in total energy is small enough to be negligible, such as less than 10 6 The system considers energy minimization complete and the iteration terminates. The total energy function is expressed as: , in, Total energy; For geometric regularity, the regularity constraint term is based on the covariance matrix of the Gaussian cell with local curvature constraint; The deviation between the coordinates of the Gaussian cell center point and the ideal centerline point corresponding to the design blueprint; This represents the overlap error between adjacent Gaussian cells, i.e., the overlap loss. These are the corresponding energy term weight coefficients.

[0046] Furthermore, when the termination condition is met, a non-maximum suppression operation is performed to screen the Gaussian cell density peaks distributed along the centerline, eliminate pseudo-extreme points caused by noise or local distortion, retain only the skeleton nodes that represent the true structural features, and output the centerline point set optimized by global consistency constraints. This not only eliminates the point cloud drift caused by structural deformation, but also accurately preserves the sharp edges and assembly features of the tower structure.

[0047] During the optimization process, a fuzzy risk assessment is simultaneously conducted, comprehensively considering the residuals of the centerline deviating from the original measurement and local curvature changes to quantify the current structural tilt uncertainty risk level. Specifically, this includes: First, the slope is extracted from the optimized centerline, and the tilt angle is output by combining it with the principal axis direction of the covariance. This tilt angle is then used to assess the interpretable tilt risk of the optimized output. The tilt vector is output as follows: , in: Indicates the angle of inclination of the centerline; indicates the slope of the structure. , , Represents the three-dimensional coordinates of the top endpoint of the structure; , , Represents the three-dimensional coordinates of the bottom endpoint of the structure.

[0048] Then, the tilt angle is transformed into a risk assessment, and probabilistic results are output. A risk membership function for the tilt angle is defined, the membership degree of the actual angle to each risk level is calculated, and warnings are issued based on the maximum membership degree to address misjudgments of critical states. The fuzzy decision is represented as: , in: This indicates the degree of risk membership; a value close to zero indicates safety. This represents the slope parameter, which controls the steepness of the membership function; The larger the scale, the sharper the decision-making boundaries; This represents the threshold parameter, which determines the critical tilt angle for risk assessment; This indicates the angle of inclination of the centerline and the slope of the structure.

[0049] It should be noted that the tilt vector output and fuzzy decision are used as physical verification and operational decisions for model optimization. The optimized centerline coordinates directly determine the spatial pose accuracy of the reconstructed model. The angle and confidence interval provided by the tilt vector are used to inversely correct the covariance principal axis direction of the Gaussian cells, eliminating structural distortions in the reconstructed model. The qualitative risk conclusions output by the fuzzy decision are embedded as high-order semantic information into the attribute table of the reconstructed model, realizing a three-dimensional integrated representation of "geometric model-physical state-risk level," providing an intuitive decision-making basis for operation and maintenance.

[0050] Specifically, in step S4, based on the optimized set of center points, and combined with the normal vectors obtained from the decomposition of the covariance matrix of each node and the local surface features, the surface reconstruction constructs a triangular manifold mesh to restore the three-dimensional surface morphology of the transmission tower. Finally, a transmission tower tilt model with optimized details is output. This model not only achieves accurate global pose correction, but also preserves the geometric details of key connection parts through local refinement, which can provide high-fidelity data support for subsequent transmission tower stability assessment.

[0051] Compared with existing technologies, this embodiment provides a detailed optimization method for transmission tower tilt models. It performs Gaussianization modeling on point clouds extracted from tilted images of the transmission tower and utilizes a deep learning network to extract geometric feature sequences of Gaussian cells, outputting component-level deformation risk probabilities and providing prior weights for the Gaussian parameter physical optimization process. Based on these prior weights, an adaptive update mechanism for Gaussian parameters is established, which inversely modulates the covariance contraction strength of the Gaussian cells and the restraint strength of the high-confidence centerline anchor points. This strongly pulls back macroscopic drift in healthy regions and actively loosens the restraints in deformed regions to preserve true deformation features, ultimately achieving high-precision, noise-resistant tilt state recognition and further obtaining a high-precision transmission tower tilt model.

[0052] Those skilled in the art will understand that all or part of the processes of the methods described in the above embodiments can be implemented by a computer program instructing related hardware, and the program can be stored in a computer-readable storage medium. The computer-readable storage medium may be a disk, optical disk, read-only memory, or random access memory, etc.

[0053] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for optimizing the details of a transmission tower tilt model, characterized in that, Includes the following steps: The original point cloud of the transmission tower is generated from multi-view images of the transmission tower, and a gridded effective point cloud of the tower body is extracted from it. Geometric feature sequences are extracted from the effective point cloud of the tower body according to the spatial position of the grid. The geometric feature sequence is input into a pre-trained classification network to obtain the deformation probability of local components of the tower. Gaussianization modeling is performed on the effective point cloud of the tower to obtain several Gaussian cells and corresponding Gaussian parameters. Based on the Gaussian parameters, the Gaussian overlap of adjacent Gaussian cells is calculated in different regions. Based on the Gaussian overlap and the deformation probability of local components, the Gaussian parameters are adaptively updated to obtain the optimized centerline point set of the tower. Based on the optimized tower centerline point set, solid modeling is performed to obtain a detailed optimized transmission tower tilt model.

2. The method according to claim 1, characterized in that, The Gaussian parameters include the center point coordinates and the covariance matrix; adaptively updating the Gaussian parameters includes: The covariance matrix of each Gaussian cell is adaptively updated based on the Gaussian overlap and the deformation probability of local components. Based on the position of the center point of each Gaussian cell to be updated in the tower structure, the differentiated neighborhood fusion range corresponding to each Gaussian cell to be updated is determined; based on the distances between each other Gaussian cell within the neighborhood fusion range and the corresponding Gaussian cell to be updated, combined with the deformation probability of local components, the dynamic weights corresponding to each other Gaussian cell are determined; based on the center point coordinates of each other Gaussian cell and the corresponding dynamic weights, spatial weighting is performed to obtain the updated center point coordinates corresponding to each Gaussian cell to be updated.

3. The method according to claim 2, characterized in that, The dynamic weights decrease exponentially with respect to the square of the distances to other Gaussian cells within the neighborhood fusion range; the dynamic weights of each Gaussian cell are determined based on the following constraints: Based on the deformation probability of local components, the confidence level of the center point corresponding to the coordinates of the center point of each Gaussian cell is determined; if the Gaussian cell to be updated is located in the edge region, the neighborhood fusion range is expanded, and the dynamic weights of other Gaussian cells within the neighborhood fusion range are proportional to the confidence level of the corresponding center point. If the Gaussian cell to be updated is located in a smooth region, then the neighborhood fusion range is shrunk.

4. The method according to claim 2, characterized in that, Adaptively update the covariance matrix corresponding to each Gaussian cell, including: Calculate the local curvature of the region where each Gaussian cell is located; The update weights corresponding to the covariance matrix of each Gaussian cell are preset. The update weights are directly proportional to the Gaussian overlap of each Gaussian cell and inversely proportional to the corresponding local curvature. The covariance matrix of each Gaussian cell is expanded or reduced based on the corresponding update weights.

5. The method according to any one of claims 1-4, characterized in that, The optimized set of tower centerline points is obtained based on the following process: Based on the Gaussian parameters, the critical contact area and non-critical contact area of ​​the tower structure are identified, and the Gaussian overlap of the corresponding areas is calculated. An overlap loss is constructed based on the Gaussian overlap degree. The optimized gradient of each center point is calculated based on the overlap loss. The coordinates of the center points of each Gaussian cell are iteratively and adaptively updated along the direction of each negative optimized gradient. In each iteration, the slope of the center line is calculated based on the updated center point coordinates. The covariance matrix is ​​adaptively updated by the rate of change of the center line slope. The total energy function is calculated based on the updated center point coordinates and the covariance matrix. The iteration process is completed by minimizing the total energy function to obtain the optimal Gaussian parameters of each Gaussian cell. The optimized tower center line point set is obtained from the center point coordinates of the optimal Gaussian parameters.

6. The method according to claim 5, characterized in that, Based on various Gaussian parameters, the critical and non-critical contact areas of the tower body are identified, including: Traverse all Gaussian cells and calculate the Euclidean distance between the center points of any two Gaussian cells to obtain the corresponding spatial distance. If the spatial distance is less than a preset distance threshold and the principal direction of the corresponding covariance matrix points to the other, then the two Gaussian cells are identified as critical contact areas; otherwise, the two Gaussian cells are identified as non-critical contact areas.

7. The method according to claim 5, characterized in that, Calculate the Gaussian overlap of the corresponding region, including: Based on the center point coordinates and covariance matrix of each Gaussian cell in the critical contact area, the weighted sum of the overlap volume of the Gaussian probability density function and the Bach coefficient is calculated to obtain the Gaussian overlap of the critical contact area. Based on the center point coordinates and covariance matrix of each Gaussian cell in the non-critical contact area, the weighted sum of the intersection volume of the envelope boxes and the spatial distance of all Gaussian cells is calculated to obtain the Gaussian overlap of the non-critical contact area.

8. The method according to claim 5, characterized in that, The overlap loss is expressed as: ， in, The overlap loss value is used to measure the degree of overlap between all Gaussian cells; Indicates the first Gaussian cells and the Gaussian cells Gaussian overlap; This indicates the preset target overlap threshold.

9. The method according to claim 5, characterized in that, The total energy function is expressed as: ， in, This represents the total energy value. For the regularity constraint term of the covariance matrix of Gaussian cells based on local curvature constraints; The deviation between the coordinates of the Gaussian cell center point and the center line point of the design blueprint; This represents the overlap error between adjacent Gaussian cells, i.e., the overlap loss. These are the corresponding energy term weight coefficients.

10. The method according to claim 1, characterized in that, The classification network is an LSTM classification network; the deformation probabilities of local components of the tower are obtained, including: The geometric feature sequence is decomposed into multiple IMF components using variational mode decomposition. Singular value decomposition is then performed on each IMF component, and a preset number of singular values ​​are extracted to form a feature vector, thus obtaining the singular value sequence corresponding to each IMF component. Each singular value sequence is input into an LSTM classification network, and the deformation probability of each local component of the tower is obtained from the output.

Images