Tower displacement detection method and device based on laser point cloud tower top feature extraction

Abstract

This invention provides a method and device for tower displacement detection based on laser point cloud tower top feature extraction. The method includes: performing binary segmentation on the acquired tower laser point cloud to achieve precise separation of the tower from the transmission line; extracting point cloud data from the tower top region from the segmented tower point cloud and removing interference from overhead ground wire point cloud to obtain a pure rigid structure point cloud at the tower top; extracting feature anchor points from the pure rigid structure point cloud at the tower top, and combining this with the symmetry characteristics of the tower structure to obtain stable feature anchor points that are spatially unique, distributed on the same horizontal reference plane, and matched with the symmetrical structure of the tower; processing the tower top feature anchor points from different measurements to calculate the spatial displacement of the tower top and the horizontal torsional deformation angle of the tower; classifying the risk level of the tower structure and generating tower displacement detection results. This invention integrates the use of tower top symmetry and other features for feature anchor point refinement, achieving centimeter-level multi-dimensional monitoring, and is particularly suitable for ultra-high voltage transmission lines.

Description

Technical Field

[0001] This invention relates to the field of abnormal detection of transmission line tower structures, specifically to the field of ultra-high voltage transmission line detection, and particularly to a tower displacement detection technology based on laser point cloud tower top feature extraction. Background Technology

[0002] Ultra-high voltage (UHV) transmission lines refer to AC or DC transmission lines with voltage levels of 1000kV and above, or ±800kV and above. As the core backbone of the national energy transmission system, UHV lines possess significant technical advantages such as large transmission capacity, long transmission distance, low energy loss, and small footprint. Their single-circuit transmission capacity is typically more than three times that of traditional 500kV lines, enabling large-scale dispatch of power resources across regions and provinces, spanning thousands of kilometers.

[0003] Ultra-high voltage (UHV) power transmission towers, as the physical carriers supporting these power arteries, are typically much larger than ordinary towers, with taller towers and bearing enormous mechanical tension from multi-branched conductors. Under these extreme conditions, any slight instability in the tower structure can trigger a chain reaction. Traditional inspection methods are often insufficient to address the ultra-high and ultra-heavy characteristics of UHV towers, especially in complex geological environments and extreme weather conditions. Even a slight displacement or tilt of 5-10 centimeters at the top of the tower, amplified over a span of hundreds of meters, can lead to insufficient conductor-to-ground clearance, flashover discharge, or even catastrophic accidents such as line breakage and tower collapse.

[0004] As the core supporting structure of ultra-high voltage (UHV) power transmission projects, the operational status of transmission towers directly affects the safety and stability of cross-regional energy transmission. UHV towers are typically characterized by their extreme height, complex structure, and enormous loads, and are mostly located in outdoor environments with variable geological conditions and frequent extreme weather events. During long-term operation, factors such as foundation settlement, wind loads, and uneven icing can easily cause minute spatial displacements at the top of the towers. Due to the large span and high tension of UHV lines, even a displacement of 5-10 centimeters at the top of the tower may indicate structural fatigue or a potential for tower collapse. Therefore, achieving routine and high-precision monitoring of the displacement at the top of UHV towers is a critical issue that urgently needs to be addressed in the field of power operation and maintenance.

[0005] 1) Currently, using drones equipped with LiDAR to collect 3D point clouds has become the mainstream inspection method. However, traditional technologies face a serious challenge in separating towers and lines when processing UHV inspection point cloud data. In the original point cloud data, tower entities, multi-split conductors, and overhead ground wires are highly intertwined in spatial location, and due to the enormous scale of UHV equipment, the amount of point cloud data grows exponentially. Existing general semantic segmentation models mostly adopt a multi-classification architecture, attempting to simultaneously identify multiple elements such as towers, conductors, insulators, shielding rings, and surrounding vegetation. However, in engineering practice, the finer the classification granularity, the lower the model's ability to distinguish similar features. Especially at the tower head attachment point, complex hardware point clouds are often misclassified, leading to a significant decrease in the model's generalization ability and accuracy in complex environments, making it impossible to provide a clean data foundation for subsequent displacement calculations.

[0006] For the quantitative calculation of tower displacement, existing technologies mostly employ full-tower point cloud matching algorithms, which are highly susceptible to data quality issues. Firstly, the overhead ground wire is directly connected to the support at the very top of the tower, and its wind-driven swaying characteristics are completely different from the displacement characteristics of the tower's rigid structure. If ground wire interference cannot be completely eliminated before calculation, the flexible noise of the ground wire will directly contaminate the displacement vector, leading to serious false alarms. Secondly, due to the systematic errors and random drift of the UAV positioning system during different inspections, direct spatial coordinate comparison makes it difficult to distinguish between measurement errors and physical displacement. There is a lack of a closed-loop processing flow that can automatically extract and robustly compare key feature areas at the tower top.

[0007] In existing point cloud processing technologies for power transmission towers, researchers mostly employ traditional machine learning algorithms based on geometric attributes for target segmentation and extraction. Clustering methods, such as DBSCAN (density-based clustering with noise) and region growing algorithms, are widely used to extract tower point clouds from complex background noise. The core logic of these methods is to aggregate points using Euclidean distance or local point cloud density distribution, identifying spatially continuous point sets as tower entities. However, when dealing with ultra-large-scale, high-density laser point cloud data like that of ultra-high voltage power transmission lines, these purely geometric statistical clustering methods exhibit significant vulnerability.

[0008] 2) Traditional clustering algorithms struggle to accurately pinpoint key structural feature points on towers. UHV towers are not solid entities but truss structures composed of numerous angle steel beams. During high-speed scanning, the laser beam's trajectory on the tower's steel frame exhibits significant randomness. This means that during inspections at different times, angles, or ambient lighting conditions, the physical locations detected by the laser point cloud often do not overlap. Since clustering algorithms like DBSCAN treat the point cloud merely as a disordered cluster of points, they cannot understand the tower's topology semantically, nor can they extract uniquely meaningful geometric feature points from the chaotic steel frame point cloud.

[0009] These methods lack stable "feature anchor points," directly leading to distortion in displacement comparison results. When comparing the state of towers at different inspection times, relying solely on the centroids or edges of clustered points for matching introduces significant computational errors. Due to inconsistencies in point cloud density between two inspections and differences in sampling point locations caused by UAV track offsets, even if the tower itself has not undergone any physical displacement, the cluster centers obtained by the clustering algorithm will experience centimeter-level spurious drifts due to the random distribution of sampling points. This "computational noise" caused by algorithmic flaws often masks the tower's true, minute deformations.

[0010] The difficulty in anchoring feature points renders existing displacement detection methods inadequate for monitoring accuracy requirements of 5-10 cm. Without structural semantic constraints, traditional comparison algorithms cannot distinguish whether changes in point cloud positions originate from sensor system errors or from actual tower structural displacement. Therefore, accurately identifying the tower structure from chaotic, sparse, and random raw point clouds using advanced deep learning techniques, and extracting stable feature primitives based on geometric constraints to eliminate interference from sampling randomness, has become a critical technological bottleneck that urgently needs to be overcome in the field of UHV precision inspection.

[0011] 3) With the rapid development of deep learning technology, semantic segmentation models based on convolutional neural networks or Transformer architectures have demonstrated significant advantages in the field of power point cloud processing. Compared with traditional geometric clustering algorithms, deep learning can achieve extremely high-precision separation of "towers" and "lines" by learning high-order features of UHV towers and power lines in terms of spatial topology, local normal vectors, and reflection intensity. This semantic-level recognition capability provides the technical prerequisite for extracting pure tower skeletons from messy inspection data. However, although existing deep learning models are constantly improving in classification accuracy, how to deeply couple the segmentation results with the structural health monitoring of towers (especially displacement detection) remains a relatively unexplored area in current research.

[0012] Patent CN119861381BA discloses a method for detecting local deformation of towers based on laser point clouds from different periods. This method automatically identifies the first transverse diaphragm of the tower through layered slicing and employs Super... The 4PCS and ICP algorithms are used for multi-stage point cloud registration to extract the displacement of key nodes to assess structural anomalies. This method achieves automated registration of multi-stage point clouds and extraction of key node displacements, which is of positive significance in the field of deformation detection. However, this method still has the following areas for improvement: First, it uses the first transverse diaphragm as the registration reference and does not focus on the tower top region with the highest displacement sensitivity; second, it removes conductors and ground wires using general point cloud classification technology, but lacks a special processing mechanism for the flexible swing characteristics of overhead ground wires; third, the extracted key node coordinates are greatly affected by sampling randomness, and it only calculates displacement, which cannot effectively distinguish between tilt and torsional anomalies of the tower.

[0013] In the mechanical structural analysis of ultra-high voltage (UHV) power transmission towers, the tower top region is the core area reflecting the overall structural stability. Based on cantilever beam structures and the mechanical amplification effect, when the tower foundation experiences slight settlement or the tower body tilts slightly, due to the ultra-high scale of UHV towers (often hundreds of meters), the minute angular displacement at the base will produce the most significant spatial linear displacement at the tower top. Therefore, the coordinate offset at the tower top is not only the most intuitive manifestation of tower displacement but also the area with the highest detection sensitivity. However, existing research often focuses on the overall matching of the entire tower structure or the monitoring of the tower base, neglecting the core value of the tower top as the displacement monitoring center. This blind selection of the monitoring area leads to a situation where, when processing large-scale point clouds, a large number of tower body point clouds that are not representative of displacement dilute the feature weight of the core region, limiting further improvements in detection accuracy.

[0014] In summary, existing technologies still face the following technical challenges in the field of UHV tower displacement detection: First, there is a lack of refined feature extraction methods specifically for the tower top area, which fails to fully utilize the mechanical amplification effect to improve detection sensitivity; second, there is a lack of a quantitative elimination mechanism for flexible interference from overhead ground wires, making it difficult to ensure the purity of displacement calculation; and third, the extraction of feature anchor points is greatly affected by the randomness of sampling, and the displacement solution dimension is singular, making it impossible to comprehensively assess the combined tilt and torsional deformation state of the tower. Summary of the Invention

[0015] This invention aims to solve the technical problems existing in the laser point cloud displacement detection technology for UHV towers, such as insufficient tower-line separation accuracy, difficulty in eliminating flexible interference from overhead ground wires, instability of feature anchor points leading to displacement calculation distortion, and the inability to distinguish between tilt and torsional anomalies due to the single dimension of displacement calculation. It achieves accurate tower-line separation through binary segmentation, eliminates ground wire interference by combining anisotropic analysis and connectivity constraints, and establishes stable feature anchor points by using latitude and longitude extreme value search, horizontal plane elevation filtering, and symmetric topology matching. Ultimately, it achieves centimeter-level, multi-dimensional accurate monitoring of tower displacement and torsional state.

[0016] The present invention provides a method for tower displacement detection based on laser point cloud tower top feature extraction, comprising: Binary segmentation processing is performed on the collected laser point cloud of the tower to achieve precise separation of the tower and the line; Extract the point cloud data of the tower top region from the segmented tower point cloud, and remove the interference of the overhead ground wire point cloud to obtain the pure rigid structure point cloud of the tower top. Feature anchor points are extracted from the point cloud of the pure rigid structure at the top of the tower. Combined with the symmetry features of the tower structure, stable feature anchor points with spatial uniqueness, distributed on the same horizontal reference plane and matching the symmetry of the tower structure are obtained. The characteristic anchor points at the top of the tower are processed for different measurements, and the spatial displacement of the top of the tower and the horizontal torsional deformation angle of the tower are calculated. Based on the calculated displacement and torsion angle, the risk level of the tower structure is classified, and tower displacement detection results are generated.

[0017] Furthermore, the binary segmentation process of the collected tower laser point cloud adopts a layered Transformer architecture binary laser point cloud segmentation model. It captures the global vertical topological dependencies of the tower through a multi-head self-attention mechanism. After local feature extraction, global topology learning, multi-scale feature fusion and classification generation, the binary classification results of the tower and line and the point set mask are output to achieve tower-line binary segmentation.

[0018] Furthermore, the point cloud data of the tower top region is extracted from the segmented tower point cloud, the tower top reference elevation is locked by the quantile statistical method, noise points higher than the reference elevation are removed, and the point cloud data of the tower top region is extracted by backtracking down from the tower top reference elevation.

[0019] Furthermore, the process involves removing interference from overhead ground wire point clouds, identifying candidate ground wire points with linear characteristics by calculating the linear significance index of the point cloud, performing connected component analysis based on Euclidean distance on the remaining point cloud, and obtaining a pure point cloud containing only the rigid steel structure at the top of the tower after secondary correction.

[0020] Furthermore, the feature anchor point extraction for the pure rigid structure point cloud at the top of the tower adopts an anchor point refinement strategy that integrates latitude and longitude extreme value search, horizontal plane elevation filtering, and symmetry topology matching. First, the edge points of latitude and longitude extreme values ​​are locked, then the point set of the same horizontal reference plane is selected, and finally, the point cluster is divided into symmetrical regions according to the symmetrical structure of the tower and the centroid of the region is extracted as the final stable feature anchor point.

[0021] Furthermore, the tower top feature anchor points of different measurements are processed, and the extreme point with the largest latitude and longitude among the feature anchor points is used as the reference origin. The geodetic coordinate system is converted into a local rectangular coordinate system with the tower center as the reference, eliminating the global drift caused by the positioning system error of different voyages. Then, the mean coordinates of the feature anchor points under different measurements are calculated respectively, which are used as the geometric center of the tower top for the corresponding measurement.

[0022] Furthermore, the calculation yields the spatial displacement at the top of the tower and the horizontal torsional deformation angle of the tower. The spatial displacement is obtained by calculating the Euclidean distance between the geometric centers of the tower top at different measurements. Principal component analysis or least squares method is used to fit the major axis direction vector of the crossarm at the top of the tower at different measurements. After projecting the vector onto the horizontal plane, the included angle is calculated to obtain the horizontal torsional deformation angle. The risk level of the tower structure is classified by setting three threshold levels: health, attention, and warning, based on the displacement. Combined with the judgment result of the horizontal torsional deformation angle, the abnormal state of the tower structure is automatically classified.

[0023] On the other hand, the present invention provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the tower displacement detection method based on laser point cloud tower top feature extraction as described above.

[0024] On the other hand, the present invention provides a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the tower displacement detection method based on laser point cloud tower top feature extraction as described above.

[0025] On the other hand, the present invention provides a computer program product, including a computer program that, when executed by a processor, implements the tower displacement detection method based on laser point cloud tower top feature extraction as described above.

[0026] This invention proposes a binary laser point cloud segmentation method for towers and other elements. It improves segmentation accuracy by combining the long-distance dependency modeling capability of Transformer, and is supplemented by fine-grained truncation of tower top window and geometric denoising. This solves the technical problems of tower line mixing, ground wire interference and insufficient displacement calculation accuracy in UHV inspection, and provides reliable technical support for the refined operation and maintenance of UHV towers.

[0027] Compared with the prior art, the advantages of the present invention are: First, the tower-line separation accuracy is high. This invention constructs a binary laser point cloud segmentation model based on a stacked Transformer architecture. It captures the global vertical topological dependencies of UHV towers at the 100-meter scale through a multi-head self-attention mechanism, and accurately segments the towers and power lines as two heterogeneous targets. This overcomes the problem of low recognition in complex areas of traditional multi-classification models and provides a clean tower point cloud foundation for displacement detection.

[0028] Second, ground wire interference is thoroughly eliminated. This invention proposes a quantitative elimination method based on local anisotropy analysis, targeting the linear geometric characteristics of overhead ground wires. By calculating the linear significance index L=(λ1-λ2) / λ1 and combining it with secondary correction of connected components, it can accurately identify and eliminate ground wire point clouds, effectively avoiding misjudgments of displacement caused by the flexible swaying of ground wires, and significantly improving the reliability of displacement calculation.

[0029] Third, the feature anchor points are stable. This invention innovatively proposes an anchor point refinement strategy that integrates latitude and longitude extreme value search, horizontal plane elevation filtering, and symmetric topology matching. This method ensures the spatial uniqueness of anchor points by locking latitude and longitude extreme points, guarantees that anchor points are located on the same horizontal reference plane through elevation filtering, and divides the region through symmetric matching and takes the centroid as the final anchor point. This effectively eliminates the computational noise caused by sampling randomness and achieves stable feature extraction at the centimeter level.

[0030] Fourth, the displacement calculation is comprehensive. This invention not only calculates the geometric center displacement of multiple characteristic anchor points, but also fits the crossarm's major axis direction vector through principal component analysis to calculate its rotation angle in the horizontal plane. This allows for the simultaneous assessment of the tower's overall displacement, tilt, and torsional deformation, providing a more comprehensive basis for structural health diagnosis.

[0031] Fifth, high detection sensitivity. This invention focuses on feature extraction in a local window area at the top of the tower, making full use of the mechanical amplification effect of the cantilever beam structure. When the tower foundation experiences slight settlement or the tower body tilts slightly, the tiny angular displacement at the bottom will be amplified into a significant spatial linear displacement at the top of the ultra-high tower, thus achieving higher sensitivity displacement detection.

[0032] Sixth, the risk levels are clearly defined. This invention sets three risk thresholds: healthy (<1cm), alert (<5cm), and warning (≥5cm). Combined with rotation angle determination, it realizes automated graded early warning of abnormal states of tower structures, providing an intuitive and quantifiable basis for power grid operation and maintenance decisions. Attached Figure Description

[0033] Figure 1 This is a flowchart of the ultra-high voltage tower displacement detection method of the present invention.

[0034] Figure 2 This is a structural diagram of a binary classification point cloud segmentation model according to an embodiment of the present invention.

[0035] Figure 3 This is a schematic diagram of a three-dimensional point set during the extraction process of the tower top region in an embodiment of the present invention.

[0036] Figure 4 This is a schematic diagram illustrating the principle of the extraction process for the top region of the tower in an embodiment of the present invention.

[0037] Figure 5 This is a schematic diagram of the characteristic region at the top of the tower according to an embodiment of the present invention.

[0038] Figure 6 This is a schematic diagram illustrating the calculation of tower top displacement and skewness in an embodiment of the present invention. Detailed Implementation

[0039] The following will further explain the concept, specific structure and technical effects of the present invention in conjunction with the accompanying drawings and embodiments, so as to fully understand the purpose, features and effects of the present invention.

[0040] Example 1 An embodiment provides a tower displacement detection method based on laser point cloud tower top feature extraction, including: 1) Perform binary segmentation on the collected laser point cloud of the tower to achieve precise separation of the tower and the line; 2) Extract the point cloud data of the tower top region from the segmented tower point cloud, and remove the interference from the overhead ground wire point cloud to obtain the pure rigid structure point cloud of the tower top; the pure rigid structure point cloud of the tower top is the point cloud set of the tower top rigid steel structure after removing the interference from non-rigid structure point clouds such as overhead ground wire and environmental noise from the segmented tower top region point cloud. This point cloud set can accurately reflect the spatial position characteristics of the rigid structure of the tower top, and eliminates the interference caused by the swaying of flexible structures and random noise, providing a stable and pure point cloud data foundation for subsequent feature anchor point extraction and displacement calculation.

[0041] The extraction of point cloud data for the top region of the tower from the segmented tower point cloud can be further achieved by using quantile statistics to determine the reference elevation of the tower top, removing noise points higher than the reference elevation, and then using the reference elevation of the tower top as a basis to backtrack downwards to extract the point cloud data that forms the top region of the tower.

[0042] Further, the removal of overhead ground wire point cloud interference can be achieved by calculating the linear significance index of the point cloud to identify ground wire candidate points with linear characteristics, and then performing connected component analysis based on Euclidean distance on the remaining point cloud. After secondary correction, a pure point cloud containing only the rigid steel structure at the top of the tower is obtained.

[0043] 3) Extract feature anchor points from the point cloud of the pure rigid structure at the top of the tower, and combine them with the symmetry features of the tower structure to obtain stable feature anchor points that are spatially unique, distributed on the same horizontal reference plane, and matched with the symmetrical structure of the tower. Furthermore, an anchor point refinement strategy that integrates latitude and longitude extreme value search, horizontal plane elevation filtering, and symmetry topology matching can be adopted. First, the edge points of latitude and longitude extreme values ​​are locked, then the point set of the same horizontal reference plane is selected, and finally the point cluster is divided into symmetrical regions based on the symmetrical structure of the tower and the centroid of the region is extracted as the final stable feature anchor point.

[0044] 4) Process the characteristic anchor points at the top of the tower for different measurements, and calculate the spatial displacement of the top of the tower and the horizontal torsional deformation angle of the tower. Furthermore, the extreme point with the largest latitude and longitude among the feature anchor points can be used as the reference origin to transform the geodetic coordinate system into a local rectangular coordinate system with the tower center as the reference, thereby eliminating the global drift caused by positioning system errors in different voyages. Then, the mean coordinates of the feature anchor points under different measurements are calculated respectively, which serve as the geometric center of the tower top for the corresponding measurement.

[0045] The spatial displacement can be obtained by calculating the Euclidean distance between the geometric centers of the tower tops at different measurement cycles. The major axis direction vectors of the crossarms at the tops of different measurement cycles can be fitted using principal component analysis or least squares method. After projecting the vectors onto the horizontal plane, the included angle can be calculated to obtain the horizontal torsional deformation angle.

[0046] 5) Based on the calculated displacement and torsional angle, the risk level of the tower structure is classified, and tower displacement detection results are generated. Specifically, three threshold levels—healthy, alert, and warning—can be set based on the displacement, and combined with the judgment results of the horizontal torsional deformation angle, the abnormal state of the tower structure can be automatically classified.

[0047] In practical implementation, it can include using innovative point cloud segmentation and displacement assessment methods to assess tower displacement in power facility scenarios, which is particularly suitable for ultra-high voltage transmission line detection. This mainly includes two stages: preprocessing and calculation. In the preprocessing stage, after obtaining the pure point cloud at the top of the tower, a search for the maximum latitude and longitude extrema is performed. By traversing all point coordinates, edge points with latitude and longitude extrema are identified, i.e., the extrema of these points whose latitude or longitude is the candidate point at the top of the tower. Then, elevation horizontal filtering is performed. Utilizing the characteristics of the main plane at the top of the UHV tower, a horizontal elevation tolerance is set, and point sets located on the same horizontal reference plane are selected. This step uses the design elevation of the main plane as a reference to ensure that the selected feature points have a high degree of consistency in the vertical direction, thereby eliminating interference from non-structural protrusions. Utilizing the structural features of the top of the UHV tower, with the tower axis centerline as a reference, the identified corner point clusters are divided into multiple symmetrical regions. This step refines the complex point cloud of the tower top steel structure into multiple key anchor points with physical meaning, stable positions, and distributed on the same horizontal plane. This not only greatly reduces the amount of data for displacement comparison but also eliminates computational fluctuations caused by sampling randomness through the dual anchoring of corner features and extreme latitude and longitude.

[0048] During the displacement calculation phase, multi-measurement feature point cloud alignment and normalization are performed. First, the centroid point sets of the feature regions from the baseline and current measurements are retrieved. Using the extreme points with the largest latitude and longitude in each set as the reference origin, the geodetic coordinate system is converted to a local Cartesian coordinate system based on the tower center, effectively eliminating global drift caused by positioning system errors from different voyages. Subsequently, the geometric center of the tower top is extracted, and the mean coordinates of the centroids of multiple corresponding feature regions under both measurements are calculated.

[0049] The actual spatial displacement at the top of the tower is obtained by comparing the Euclidean distance between the geometric centers of the two measurements. In the assessment of structural skew angle and abnormal states, principal component analysis or least squares method is used to perform linear fitting on multi-region feature points for each measurement. The degree of horizontal torsional deformation of the tower is quantified by projecting the major axis vectors of the two measurements onto the horizontal plane and calculating their included angle.

[0050] Example 2 This invention proposes a displacement detection method for ultra-high voltage (UHV) towers based on deep learning-based binary segmentation and geometric topological constraints. First, a Transformer architecture model is used to perform binary segmentation of the tower and power line in the inspection point cloud, solving the problem of target interleaving in complex environments. Then, a local area (e.g., 1 to 2 meters) at the top of the tower is automatically extracted, and ground wire interference is removed through local anisotropy analysis to obtain a clean tower top skeleton.

[0051] Based on this, stable feature anchor points are locked by searching for extreme latitude and longitude points and filtering on the horizontal plane, and a local coordinate system is established to eliminate positioning drift. By comparing the centroid displacement of anchor points and the skew angle of the crossarm major axis in multiple measurements, the degree of tower deformation is quantified. Finally, based on the offset threshold, the system automatically classifies the tower into health, alert, and early warning levels and outputs maintenance decisions, achieving robust monitoring of UHV towers with centimeter-level accuracy.

[0052] Specifically, addressing the need for monitoring minute displacements of UHV transmission towers, this invention constructs a binary laser point cloud segmentation model oriented towards towers and other elements. It utilizes a multi-head self-attention mechanism within a stacked Transformer architecture to capture global vertical topological features, achieving precise tower-line separation in complex scenarios. This method differs significantly from traditional full-tower matching. It extracts key areas at the tower top through adaptive elevation trimming and eliminates overhead ground wire interference by combining local anisotropy analysis and connectivity constraints, obtaining a pure, rigid tower top skeleton. Furthermore, it proposes a feature anchor point refinement strategy, using latitude and longitude extreme value search and horizontal plane elevation filtering to lock spatially stable nodes, and eliminating the influence of sampling randomness through symmetric topological matching.

[0053] Based on the high-precision feature extraction results described above, this invention can be used for displacement vector calculation and structural skew diagnosis of multi-measurement inspection data. By calculating the geometric center offset and the crossarm major axis rotation angle, it achieves automated, graded early warning of overall tower tilt, torsional anomalies, and foundation settlement. This invention effectively solves the problems of ground wire noise interference and positioning system drift in UHV inspections, significantly improves the robustness of tower deformation monitoring, and provides accurate decision support for refined power grid operation and maintenance.

[0054] See Figure 1 The tower displacement detection method based on laser point cloud tower top feature extraction provided in the embodiment, and the ultra-high voltage tower displacement detection method based on deep learning and geometric topology constraints, mainly realizes the automated conversion from point cloud to centimeter-level anomaly monitoring results through the following six core steps.

[0055] S101, Perform binary laser point cloud segmentation In practical implementation, existing point cloud classification deep learning network structures can be used. Alternatively, the point cloud classification deep learning network structure established in this embodiment can be used as a binary laser point cloud segmentation model for towers and other elements, with better results. When performing binary segmentation processing on the collected tower laser point clouds, this embodiment uses a layered Transformer architecture binary laser point cloud segmentation model. It captures the global vertical topological dependencies of the towers through a multi-head self-attention mechanism. After local feature extraction, global topology learning, multi-scale feature fusion, and classification generation, it outputs the binary classification results of the towers and lines, as well as point set masks, to achieve tower-line binary segmentation.

[0056] See Figure 2The network structure provided in this embodiment is a four-stage deep learning architecture for target classification and segmentation tasks using LiDAR point clouds and structured tabular data. It sequentially combines point cloud feature extraction based on sampling / grouping, a stacked Transformer backbone network, multi-scale feature sampling and fusion, and head prediction based on an attention mechanism. Based on this network structure, binary segmentation is achieved through four stages: first, local features are extracted using FPS sampling and K-NN; second, the global topology is captured through a multi-layer self-attention mechanism (e.g., 12 layers); then, multi-scale features are fused to generate a query vector; finally, the binary classification results for towers and lines, along with a point set mask, are output using a cross-attention mechanism.

[0057] In one possible implementation, the specific implementation of each stage is as follows: Phase 1: Feature Extraction First, point sampling is performed, with a preferred recommendation of two-stage downsampling, processing N={8192, 2048} points. Next, local geometric information is aggregated through K-NN grouping (e.g., K=32, parameter). Subsequently, a base feature embedding layer (e.g., dimension K:{16,32}, hyperparameter) and a multilayer perceptron (MLP) (e.g., dimension K:{64, 128}, hyperparameter) map the local features. At this stage, structured metadata table information is processed and mapped into a representation vector, which is then incorporated as contextual information.

[0058] Phase 2, Transformer The preferred approach consists of a three-layer stacked Transformer backbone network, comprising three Transformer layers (denoted as Transformer Layer 1, 2, and 3, the number of layers can be adjusted in the specific implementation). Features are passed between layers through hierarchical input-output dependencies. The implementation can refer to the classic Transformer structure. Each layer contains a LiDAR Multi-Head Self-Attention Module (MHSA) and a Feedforward Network (FFN) module. In the MHSA module, the network fuses metadata context in the hidden layer dimension (hyperparameter). The MHSA module preferably uses 12 attention heads (hyperparameters). The FFN module is specified with a dimension of 24 (FFN:24, hyperparameter). Each MHSA or FFN module has layer normalization and residual connections set before and after it, respectively. The result of each self-attention calculation can be stored in a kv cache for subsequent feature extraction queries.

[0059] Phase 3, Multi-scale Sampling This stage has two branches: in the downsampling branch (i.e., MLP downsampling blocks), features are downsampled and fused from multiple resolution levels. Each downsampling block is formed by MLP combined with mean pooling operations (…). Figure 2 Pooling is represented by a circle with a plus sign to aggregate features. In practice, global queries can be generated using MLP structures corresponding to different spatial scales (input dimensions [64, 128]). Meanwhile, in the feature fusion and reduction branch (i.e., feature fusion & reduction block), feature fusion is used to align features, ensuring that the queries and HW features with cross-attention in sampling have the correct dimensional representation. The three vertical MLP layers after attention complete the function of FFN in the standard Transformer, providing deep representation learning, and finally mapping the features to the classification output.

[0060] Phase 4: Classification and Generation This process occurs within a classifier head that employs a cross-attention mechanism to refine predictions. It uses an attention mechanism to process the fused feature map (256, 128, HW) and the query features (256, 128, C). The cross-attention layer combines local and global context (HW dimension) with the query (C dimension).

[0061] Based on the above-mentioned point cloud classification deep learning network structure, in a preferred embodiment, automatic extraction of laser point cloud tower top features and displacement calculation are realized for automatic displacement detection of UHV transmission towers, including: Step 1. Multi-level local feature extraction In one possible embodiment, step 1 is implemented as follows: 1-1 First-level downsampling: Using the farthest point sampling (FPS) algorithm, the input original point cloud patch is processed into... N 1 = 8192 key points, preserving the global outline.

[0062] 1-2 Second-level downsampling: Based on the first-level downsampling, further FPS sampling is performed to obtain... N 2 = 2048 skeleton points are used to reduce the runtime complexity of subsequent Transformers.

[0063]

[0064] in, P express N The point set in 1, p i For any one of them, P select This represents the selected set of sampling points. p jLet be any one of them. ||.||2 represents two points. p i and p j The Euclidean distance between them. min represents the minimum distance, and argmax() represents the search for the distance that satisfies the condition. p i .

[0065] 1-3 Perform local neighborhood grouping, where the neighborhood search parameters are: for N With a sampling point size of 2, the K-Nearest Neighbor (K-NN) algorithm is used to search for local spatial point clusters. It is recommended to set the hyperparameter neighborhood number of points K=32.

[0066] 1-4 Feature Embedding: Parallel feature embedding of hyperparameters [64, 128] is performed to ensure that the model can simultaneously capture the geometric features of point clouds of different categories. In the example, it can simultaneously capture the features of the edge of the UHV angle steel (fine-grained) and the tower frame (coarse-grained).

[0067] 1-5 Multilayer Perceptron (MLP) Configuration: Geometric coordinates are integrated using multi-level MLP operators. x , y , z Mapping to a higher-dimensional space, with feature dimensions progressively set to {64, 128}, the linear transformation formula is as follows:

[0068] In the formula For channel i In the l The output features of the layer For the first l The output feature vector of layer -1 is also the first layer. l The input feature vector of the layer, For the weight vector, For bias terms, It is a non-linear activation function.

[0069] 1-6 Feature Concatenation: Concatenate the local features extracted from different groups to form an initial feature vector containing local geometric information.

[0070] Step 2. Stack Transformers In one possible embodiment, step 2 can be implemented in the following way: 2-1 Multi-Head Configuration: The MSHA module employs a multi-head self-attention mechanism to capture the long-distance vertical topological dependencies of UHV towers at the 100-meter scale. The scaling dot product attention formula is:

[0071] in, Q , K , V These represent the query, key, and value matrices, respectively. is the dimension scaling factor for the feature vector, Attention() is the attention function, and softmax() is the attention calculation function.

[0072] 2-2 Stacking Depth: In the MSHA module of the Transformer block, the sampling unit is set to... K =12, ensuring deep abstraction of semantic features.

[0073] 2-3 Feedforward Neural Network (FFN): Each MHSA module is followed by an FFN module. The FFN module usually consists of two fully connected layers (linear transformation) and a non-linear activation function (such as ReLU). By setting the linear transformation and the non-linear activation function, the non-linear fitting ability of the model is enhanced.

[0074] 2-4 Fusion of Absolute and Relative Positions: The position embedding information obtained in the first stage is injected into the input of the Transformer. The first normalization layer in the Transformer (i.e., the normalization layer set before the MSHA module of Transformer Layer 1) completes the position embedding of different channels and performs normalization to ensure that the model does not lose the extremely critical spatial absolute height information of the UHV tower while learning global semantics.

[0075] Step 3. Multi-scale sampling In one possible embodiment, step 3 can be implemented in the following way: 3-1 After downsampling, the feature vectors obtained from multiple MLP processing are merged through Pooling.

[0076] 3-2 Residual Connections: The MLP downsampling block also contains residual connection streams, which are also introduced as input features for Pooling.

[0077] 3-3 Class Query MLP Processing: The feature stream from the class query embedding is first processed by a 3-layer MLP within the feature fusion & reduction block for feature mapping and alignment.

[0078] 3-4 The query feature stream after MLP alignment is merged with the feature stream from the outside and a feature concatenation operation is performed to achieve the initial fusion of features of different scales and attributes.

[0079] Step 4. Classification Stage 4-1 In the classification stage, the multiple queries obtained from the query generation step in stage 3 are used as vector Q, and the fused features from stage 3 are used as vectors. F fused As a key K , value V To interact, specifically, means:

[0080] In the formula, F out The output features after semantic focusing enable the query vector to automatically focus on feature regions in the point cloud that possess pole-like semantics. `softmax()` is the attention calculation function. The key for the fused features, with the superscript T indicating vector transpose. The value of the fusion feature, Q queries The key for the query. This is the dimension scaling factor for the feature vectors (query and key vectors).

[0081] 4-2 Multiple multilayer perceptrons (MLPs) are set before and after the cross-attention layer for feature representation and feedforward deep representation learning, respectively. A 3×2 MLP block is preferred, with an input dimension of [64, 128], and the MLP structure uses the ReLU activation function.

[0082] The three vertical MLP layers preceding the cross-attention layer ensure that the queries and HW features have the correct dimensionality representation. The three vertical MLP layers following the cross-attention layer can perform the function of the FFN in the standard Transformer, providing deep representation learning and ultimately mapping the features to the classification output.

[0083] 4-3 Binary Classification Output: The probability distribution of each query belonging to a tower or line is determined by the prediction head, and the mask query results are output.

[0084] S102. Extraction of laser point cloud classification points from the top of transmission towers. This step provides a scheme for classifying and extracting point sets from the laser point cloud at the top of a power transmission tower, which is used to extract data from the top portion of the segmented point cloud.

[0085] Step 1. Tower tip benchmark elevation locking and outlier filtering 1-1 Tower-type point set extraction: Extract the point cloud set of all entities labeled as Tower from the label mask output by the binary segmentation model.

[0086] 1-2 Statistical Elevation Peak Locking: To eliminate the influence of outliers in the laser point cloud on the determination of tower height, instead of directly taking the maximum value, the quantile statistical method is used to lock the reference elevation of the tower tip. Z peak .

[0087]

[0088] in, γ The value is [0.99, 0.999]. P tower Let z be the set of point clouds of the tower. i For any point p i The elevation, Percenttile() is the percentage that satisfies the condition. γ The above are the quantile values.

[0089] 1-3 Dynamic noise removal: Noise exceeding the tower tip reference elevation will be removed. Z peak Sparse points are treated as environmental noise and directly removed to ensure that the starting point of the subsequent clipping window has physical authenticity.

[0090] S103. Overhead ground wire stripping based on geometric connectivity and linear characteristics In a preferred embodiment, a geometry-based ground wire stripping method is provided to introduce a geometric connectivity determination method within the classified tower top data to eliminate the influence of overhead ground wires.

[0091] Step 1. Overhead ground wire stripping based on connectivity determination 1-1 Preliminary Identification of Overhead Ground Wires: Within the above output point set, in addition to the steel frame of the tower, there may also be overhead ground wires spanning the tower tip. Calculations are performed using the linear distribution characteristics of the ground wires:

[0092] Where λ is the linear significance index, e 1, e 2, e 3 represents the eigenvalues ​​of the local covariance matrix. If λ > δ (the threshold is a hyperparameter, and it is recommended to set it to 0.5), then the cluster of points to be examined has a high degree of linearity and is determined to be a candidate point for the ground line.

[0093] 1-2 Connected Component Secondary Correction: Perform connected component analysis based on Euclidean distance on all point clouds in the remaining point set. Starting from any point, expand with a preset corresponding hyperparameter (e.g., 10 cm) as the radius, and heuristically access all connectable point sets as output.

[0094] 1-3 Output: The output contains only a pure point set of the rigid steel structure at the tower tip as the reference input for displacement monitoring. This point set can accurately reflect the spatial vector displacement of the tower with an accuracy of 5-10 cm, completely solving the problem of deformation misjudgment caused by the flexible swaying of the ground wire in UHV inspection.

[0095] See Figure 3 and Figure 4 Based on the binary segmentation results, the extreme value of the tower elevation is locked and 2 meters is cut back downwards. After removing the overhead ground wire through local spatial filtering and geometric shape recognition, it is possible to accurately obtain only the rigid steel structure of the tower tip.

[0096] S104, Extraction of Feature Regions from Laser Point Cloud at Tower Top In a preferred embodiment, the present invention provides a scheme for extracting the feature region of a tower top laser point cloud, which is used to perform spatial filtering on the tower top laser point cloud data.

[0097] Step 1. Primary Spatial Filtering 1-1 Search for the maximum elevation of the above output points :

[0098] 1-2 with To define the boundary, backtrack downwards by setting a height (e.g., 2 meters) to establish a three-dimensional region Φ.

[0099] Step 2. Feature point set extraction based on local anisotropy 2-1 Covariance Matrix Construction: For each sampling point within the Φ region, search for neighborhood points with radius r, and construct a local covariance matrix C:

[0100] In the formula The spatial center of the neighborhood points, k For the number of neighboring points, i This is used to label neighboring points. In practice, r can be preset with a value based on experience.

[0101] 2-2 Eigenvalue Analysis and Linear Significance Measurement: Solving for the eigenvalues ​​λ1, λ2, λ3 of matrix C, and calculating the linear significance index. :

[0102] in the formula The closer the value is to 1, the higher the probability that the point belongs to a single, extended linear structure like the ground line; if , If the value is close to 0, the probability that the point belongs to the tower top angle steel plane or the cross structure is higher.

[0103] 2-3 Maximum Latitude and Longitude Extremum Point Search: Within the extracted pure point cloud set at the top of the tower, traverse all point coordinates and identify edge points with extreme latitude and longitude values; these are the extreme points whose longitude or latitude corresponds to the candidate points at the top of the tower. The elevation corresponding to each extreme point is... .

[0104] 2-4 Elevation Horizontal Filtering: Utilizing the characteristics of the main plane at the top of the tower, such as the crossarm or the platform at the top of the tower, a horizontal elevation tolerance is set. (For example, 2 meters), filter out the set of points on the same horizontal reference plane:

[0105] in, The design elevation of the main plane ensures that multiple feature points have consistent height in the vertical direction. 2-5 Symmetry Topology Matching: Utilizing the common left-right symmetrical structural characteristics of UHV towers, the symmetry axis at the tower top (this serves as a preliminary reference and can be determined by design parameters or the center of the tower feet, such as...) Figure 6 Using the dashed lines in the diagram as a reference, the identified corner point clusters are divided into multiple symmetrical regions. Corners are the regions with the highest neighborhood information entropy, corresponding to the intersections or endpoints of angle steel. In specific implementations, the KNN method can be used to identify corner point clusters.

[0106] This step refines the complex point cloud of the tower top steel structure into multiple key anchor points that have physical meaning, stable positions, and are distributed on the same horizontal plane. This not only greatly reduces the amount of data for displacement comparison, but also eliminates non-rigid displacement interference caused by wind vibration and icing of overhead ground wires through dual anchoring of corner features and extreme latitude and longitude. This ensures that subsequent displacement calculations are performed on the stable benchmark of the rigid frame at the tower top, which is a prerequisite for achieving precise monitoring of UHV towers at the 5-10 cm level.

[0107] See Figure 5 By extracting the characteristic region at the top of the tower, this invention utilizes the symmetry of the crossarm and the extreme values ​​of latitude and longitude to lock multiple horizontally distributed stable anchor points, supporting the realization of accurate displacement monitoring.

[0108] S105. Displacement calculation of characteristic regions of towers based on laser point cloud measurements of different orders. In a preferred embodiment, the present invention provides a displacement and tilt angle calculation method based on topological constraints of multiple feature regions, which calculates the displacement and tilt angle of the tower based on the feature regions at the top of the tower in different measurement point cloud.

[0109] Step 1. Alignment and Normalization of Multi-Measurement Feature Point Clouds 1-1 Test Data Retrieval: Obtain the centroid sets of multiple feature regions extracted from the baseline test and the current test respectively.

[0110] 1-2 Geographic coordinates to local coordinates conversion: Using the extreme point with the largest latitude and longitude in each set of points as the reference origin, the geodetic coordinate system is converted into a local rectangular coordinate system with the tower center as the reference, eliminating global drift caused by GNSS system errors in different voyages.

[0111] 1-3 Extraction of the geometric center of the tower top: Calculate the mean coordinates of the centroids of multiple feature regions (one-to-one correspondence between the two measurements) under two measurements respectively. , serving as the geometric center of the tower top structure under this measurement:

[0112] in, N The number of characteristic regions of the tower. For the first k The geometric center coordinates of the i-th feature region in each measurement.

[0113] 1-4 Based on Calculate the displacement for two measurements, where displacement is defined as... and The Euclidean distance.

[0114] Step 2. Structural Skew Angle and Abnormal State Assessment 2-1 Fitting the long axis direction vector: Using principal component analysis (PCA) or the least squares method, linear fitting is performed on the points in the multi-region of each measurement to extract the long axis of symmetry vector representing the extension direction of the tower top, that is, the precise direction vector of the tower top symmetry axis, which is used to calculate the horizontal skew angle.

[0115] 2-2 Calculation of horizontal skew angle: Project the major axis vectors of the two measurements onto the horizontal XY plane and calculate the included angle.

[0116] S106. Generation of tower displacement detection results In a preferred embodiment, the present invention provides results for determining displacement detection results and generating auxiliary operation and maintenance decisions, including but not limited to: 1-1 Characterization of Rotation Angle: The tilt angle of the long axis of symmetry at the top of the tower in the horizontal plane is given to quantify the torsional deformation of the tower structure. 1-2 Deviation Ratio Verification: Compare the measured values ​​with the safety threshold, calculate the deformation deviation rate, and use it as the physical basis for determining the safety level. 1-3 Overall Tilting Diagnosis: If the displacement is significant but the tilt angle is constant, it is determined that the tower is tilted overall, mainly due to uneven settlement of the foundation. 1-4 Structural Torsion Diagnosis: If the skew angle increases significantly, it is determined to be an abnormal torsion caused by uneven stress on the top of the tower or structural loosening. 1-5 Foundation Settlement Assessment: Determine whether there is a risk of foundation settlement based on the absolute value of vertical displacement. 1-6 Key Area Comparison: Display the spatial distribution differences of characteristic corner areas in two inspections to ensure that the judgment is supported by physical topology. 1-7 Risk Level Classification: Based on the degree of deformation, it is automatically marked into three levels: healthy (the initial and current two measurements are offset by less than 1 cm), alert (the initial and current two measurements are offset by less than 5 cm), and warning (the initial and current two measurements are offset by more than 5 cm).

[0117] See Figure 6 The diagram illustrates the calculation principle of tower top displacement and skew. By comparing the characteristic anchor points of two measurements, a displacement of 12.7 cm and a rotation angle of 4.9° are quantified, thus enabling the determination of structural deformation anomalies.

[0118] To facilitate understanding of the technical effects of the present invention, the following implementation verification data of the embodiments are provided: The experimental verification of this invention is based on measured laser point cloud data of a 1000kV UHV transmission line at a UHV experimental base. The original input point cloud reached 12 million sampling points, covering UHV towers, eight-split conductors, overhead ground wires, and complex ground environments. The experimental comparison object is the currently industry-standard DBSCAN clustering method combined with the centroid displacement calculation method of the entire point cloud.

[0119] During the classification stage, the traditional DBSCAN algorithm, affected by the complexity of UHV fittings and the uneven point cloud density, cannot accurately separate the split conductors immediately adjacent to the tower head. Experiments showed that its classification bias in the tower head region was relatively large, resulting in a mean spatial measurement error of 11.23 cm for the eight extracted reference points (involving the angle steel apex, crossarm edge point, and tower top support anchor point on the -2 meter horizontal plane of the tower top). In contrast, this invention uses the Transformer binary segmentation model to accurately identify the tower body through global topology perception. Experimental results show that the spatial positions of the eight reference points extracted by this invention are extremely stable, and the measurement errors are shown in Table 1.

[0120] Table 1. Statistics of errors in two measurements of feature points and towers (error unit: cm)

[0121] In the final tower displacement calculation stage, traditional methods, due to their direct processing of the entire point cloud, cannot avoid the non-rigid displacement noise generated by the wind-driven swaying overhead ground wires. In benchmark tests where the towers did not actually experience displacement, the average pseudo-displacement (system noise) calculated by traditional methods reached 11.23 cm, meaning it cannot effectively identify the real minute deformations within the 5-10 cm range. This invention eliminates drift by extracting characteristic anchor points at the tower top and establishing a local Cartesian coordinate system. Experimental data shows that the overall tower offset error calculated by this invention is only 2.73 cm, with a comprehensive accuracy improvement of 75.7% compared to traditional methods. Within the 5-10 cm UHV core monitoring range, this invention exhibits extremely high recognition sensitivity and linearity.

[0122] In terms of computational efficiency, this invention demonstrates significant engineering advantages. The traditional DBSCAN full-set method requires intensive search and iterative calculations on all 12 million points, resulting in lengthy computation times. However, this invention employs a semantically guided local window pruning strategy, locking onto key feature regions at the tower top, so the number of point clouds actually involved in the core calculations accounts for only about 12% of the original data. This targeted data dimensionality reduction makes the algorithm's processing speed far faster than traditional methods. While maintaining centimeter-level monitoring accuracy, it significantly reduces the consumption of hardware computing resources, better meeting the needs of rapid inspection and real-time auxiliary decision-making for large-scale ultra-high voltage lines.

[0123] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

[0124] In specific implementation, the method proposed in the technical solution of this invention can be automatically executed by those skilled in the art using computer software technology. System devices for implementing the method, such as computer-readable storage media storing the corresponding computer program of the technical solution of this invention and computer equipment including the computer program running the corresponding computer program, should also be within the protection scope of this invention.

[0125] The apparatus provided by the present invention is described below. The apparatus described below can be referred to in correspondence with the tower displacement detection method based on laser point cloud tower top feature extraction described above.

[0126] In another embodiment, the present invention provides an electronic device that may include: a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other via the communication bus. The processor can call logical instructions in the memory to execute a tower displacement detection method based on laser point cloud tower top feature extraction.

[0127] Furthermore, the logical instructions in the aforementioned memory can be implemented as software functional units and sold or used as independent products, and can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0128] In another embodiment, the present invention also provides a computer program product, the computer program product including a computer program, the computer program being stored on a non-transitory computer-readable storage medium, and when the computer program is executed by a processor, the computer is able to execute the software processing part of the tower displacement detection method based on laser point cloud tower top feature extraction provided by the above methods.

[0129] In another embodiment, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the software processing portion of the tower displacement detection method based on laser point cloud tower top feature extraction provided by the above methods.

[0130] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0131] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0132] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for detecting tower displacement based on laser point cloud tower top feature extraction, characterized in that, include: Binary segmentation processing is performed on the collected laser point cloud of the tower to achieve precise separation of the tower and the line; Extract the point cloud data of the tower top region from the segmented tower point cloud, and remove the interference of the overhead ground wire point cloud to obtain the pure rigid structure point cloud of the tower top. Feature anchor points are extracted from the point cloud of the pure rigid structure at the top of the tower. Combined with the symmetry features of the tower structure, stable feature anchor points with spatial uniqueness, distributed on the same horizontal reference plane and matching the symmetry of the tower structure are obtained. The characteristic anchor points at the top of the tower are processed for different measurements, and the spatial displacement of the top of the tower and the horizontal torsional deformation angle of the tower are calculated. Based on the calculated displacement and torsion angle, the risk level of the tower structure is classified, and tower displacement detection results are generated.

2. The tower displacement detection method based on laser point cloud tower top feature extraction according to claim 1, characterized in that: The acquired tower laser point cloud is processed into binary segmentation using a layered Transformer architecture binary laser point cloud segmentation model. The model captures the global vertical topological dependencies of the tower through a multi-head self-attention mechanism. After local feature extraction, global topology learning, multi-scale feature fusion, and classification generation, the binary classification results of the tower and line and the point set mask are output, thus realizing the binary segmentation of the tower and line.

3. The tower displacement detection method based on laser point cloud tower top feature extraction according to claim 1, characterized in that: The point cloud data of the tower top region is extracted from the segmented tower point cloud. The tower top reference elevation is locked by the quantile statistical method. After removing noise points higher than the reference elevation, the tower top reference elevation is used as the basis to backtrack downwards to extract the point cloud data that forms the tower top region.

4. The tower displacement detection method based on laser point cloud tower top feature extraction according to claim 1, characterized in that: The process involves removing interference from overhead ground wire point clouds, identifying candidate ground wire points with linear characteristics by calculating the linear significance index of the point cloud, and then performing connected component analysis based on Euclidean distance on the remaining point cloud. After secondary correction, a pure point cloud containing only the rigid steel structure at the top of the tower is obtained.

5. A method for tower displacement detection based on laser point cloud tower top feature extraction according to claim 1, characterized in that: The feature anchor point extraction for the point cloud of the pure rigid structure at the top of the tower adopts an anchor point refinement strategy that integrates latitude and longitude extreme value search, horizontal plane elevation filtering and symmetry topology matching. First, the edge points of latitude and longitude extreme values ​​are locked, then the point set of the same horizontal reference plane is selected, and finally the point cluster is divided into symmetrical regions according to the symmetrical structure of the tower and the centroid of the region is extracted as the final stable feature anchor point.

6. A method for tower displacement detection based on laser point cloud tower top feature extraction according to claim 1, characterized in that: The tower top feature anchor points of different measurements are processed, and the extreme point with the largest latitude and longitude among the feature anchor points is used as the reference origin. The geodetic coordinate system is converted into a local rectangular coordinate system with the tower center as the reference, so as to eliminate the global drift caused by the positioning system error of different voyages. Then, the mean coordinates of the feature anchor points under different measurements are calculated respectively, which are used as the geometric center of the tower top of the corresponding measurement.

7. A method for tower displacement detection based on laser point cloud tower top feature extraction according to claim 1, characterized in that: The calculation yields the spatial displacement at the top of the tower and the horizontal torsional deformation angle of the tower. The spatial displacement is obtained by calculating the Euclidean distance between the geometric centers of the tower top at different measurements. Principal component analysis or least squares method is used to fit the major axis direction vector of the crossarm at the top of the tower at different measurements. After projecting the vector onto the horizontal plane, the included angle is calculated to obtain the horizontal torsional deformation angle. The risk level of the tower structure is classified by setting three threshold levels: health, attention, and warning, based on the displacement. Combined with the judgment result of the horizontal torsional deformation angle, the abnormal state of the tower structure is automatically classified.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, it implements the tower displacement detection method based on laser point cloud tower top feature extraction as described in any one of claims 1 to 7.

9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the tower displacement detection method based on laser point cloud tower top feature extraction as described in any one of claims 1 to 7.

10. A computer program product, comprising a computer program, characterized in that: When the computer program is executed by the processor, it implements the tower displacement detection method based on laser point cloud tower top feature extraction as described in any one of claims 1 to 7.

Images