A power transmission line hidden danger intelligent identification and early warning method, medium and system

Abstract

The application provides a power transmission line hidden danger intelligent identification and early warning method, medium and system, and belongs to the technical field of power transmission line detection. Three-dimensional point cloud data of a power transmission channel is collected by a mobile platform carrying a laser scanning device, a distributed external storage double-layer decoupling index is constructed to efficiently manage the channel point cloud, a multi-pulse echo full-wave time-space correlation separation algorithm is used to extract tower point cloud and cable point cloud, an artificial intelligence model, i.e. a curvature flow modulation fractal gradual reconstruction model, is used to complete the high-fidelity of incomplete tower point cloud, and an enhanced point cloud matching framework is used to accurately register the complete tower point cloud with the reference design point cloud, so that early warning is finally realized. The technical problem that incomplete point cloud cannot be completed with high fidelity due to local shielding of the tower, and further causes misjudgment of subsequent registration and deformation analysis of power transmission line point cloud data is solved.

Description

Technical Field

[0001] This invention belongs to the field of power transmission line detection technology, and specifically relates to a method, medium and system for intelligent identification and early warning of hidden dangers in power transmission lines. Background Technology

[0002] Transmission line inspection is a fundamental task for ensuring the safe operation of the power grid. Laser scanning technology has been widely applied in the field of 3D point cloud acquisition and hazard identification of transmission channels. In existing technologies, point cloud completion methods based on local features in the spatial domain are commonly used for the acquired tower point clouds. These methods include completion algorithms based on Poisson reconstruction, moving least squares, and spatial domain graph neural networks. These methods can recover basic geometric contours under conditions of high point cloud integrity. However, due to the complex structure of transmission towers, there are many geometric abrupt changes between crossarms, node plates, and main materials. In actual scanning, factors such as occlusion, limited flight altitude, and scanning angle deflection result in large-area missing points in the tower point cloud at crossarm connections, insulator attachment points, and densely packed node plate areas. The local features of the missing areas are completely broken, and the neighborhood features relied upon by the spatial domain methods cannot transmit topological information across the missing areas. Under the aforementioned operating conditions, the spatial domain local feature extraction operator can only perceive a limited neighborhood near the missing boundary, failing to acquire global manifold information such as the overall bending trend and symmetry of the tower. This results in defects such as topological breaks, surface wrinkles, and uneven point cloud distribution in the completion results in regions of high curvature geometric abrupt changes. The registration error between the completed point cloud and the baseline design point cloud increases significantly, leading to the misjudgment of normal structures as potential deformation hazards. In other words, existing technologies suffer from the technical problem that transmission line point cloud data cannot be fully and faithfully completed due to partial occlusion by the tower, thus causing misjudgments in subsequent registration and deformation analysis. Summary of the Invention

[0003] In view of this, the present invention provides a method, medium and system for intelligent identification and early warning of hidden dangers in transmission lines, which can solve the technical problem in the prior art where the point cloud data of transmission lines cannot be fully and accurately completed due to partial obstruction by towers, thus leading to misjudgment in subsequent registration and deformation analysis.

[0004] The present invention is implemented as follows: The first aspect of the present invention provides a method for intelligent identification and early warning of hidden dangers in power transmission lines, comprising the following steps: Three-dimensional point cloud data of power transmission channels are collected by a mobile platform equipped with laser scanning equipment. A distributed external storage dual-layer decoupled index is constructed. The channel is divided into geographical data units according to the span and tower location. The upper layer adopts a lightweight spatial grid index for macro topology, and the lower layer dynamically swaps in and out local point cloud files based on memory mapping technology. A multi-pulse echo full waveform spatiotemporal correlation separation algorithm is applied to the point cloud within the geographic data unit. The laser echo amplitude, pulse half-width, full width at half-height, and time series energy distribution are jointly analyzed to construct a single-point physical property scattering matrix, remove secondary scattering artifact points, and extract tower point clouds and cable point clouds. For the incomplete tower point cloud and fitting point cloud, the curvature flow modulation fractal progressive reconstruction model is input to complete the missing regions and output the complete tower point cloud. During the completion process, the spectral tensor kernel width parameter is dynamically given by the density adaptive adjustment function based on the point density estimate of the local region, the normal vector consistency index and the feature difference degree of adjacent levels. An enhanced point cloud matching framework based on feature descriptors is adopted to perform three-layer pyramid multi-resolution iterative nearest point registration with the benchmark design point cloud to obtain an accurate spatial transformation matrix. Based on the spatial transformation matrix, the tower translation vector and rotation angle are extracted. The catenary physical force equation and the spatial variation principle cable sag deformation decoupling algorithm are applied to the cable point cloud. The deformation residual is quantitatively output as the hidden danger feature. Density-based spatial clustering is performed on the over-limit points to merge the hidden danger areas. A multi-factor weighted comprehensive risk assessment model was established, extracting tower deformation parameters, cable spacing parameters, and characteristic parameters of potential hazard areas. Voltage level, ambient temperature, and wind speed correction coefficients were introduced to dynamically adjust the safe distance threshold, and the risk level was divided into four levels. An interactive three-dimensional visualization scene was constructed based on WebGL technology, generating a structured early warning report and pushing it to the inspection management system.

[0005] Specifically, the distributed external storage dual-layer decoupled index divides the power transmission channel into independent geographic data units with high cohesion and low coupling in geographic space. The upper layer is a lightweight spatial grid that stores the bounding box coordinates and file pointers of the spatial boundary of each geographic data unit. The lower layer uses memory mapping technology to directly map the disk address space of the point cloud file to the virtual address space of the process, allowing for swapping in and out as needed. This decouples the computing power of each geographic data unit and supports multi-processor parallel scheduling.

[0006] Specifically, the multi-pulse echo full waveform spatiotemporal correlation separation algorithm performs Gaussian decomposition on the full waveform digital record, fits the synthesized echo into the sum of several single-peak Gaussian components, establishes a single-point physical property scattering matrix for each scattering event, introduces spatiotemporal coherence constraints to eliminate secondary scattering artifacts, and assigns the point cloud to the corresponding physical target based on the time and amplitude of each scattering component.

[0007] In the Gaussian decomposition, the upper limit of the number of components is 6. This value is determined by performing a full waveform scan of the parallel steel cable in a laboratory environment and using the Akaike Information Criterion as the model selection basis.

[0008] Specifically, the curvature flow modulation fractal progressive reconstruction model is composed of a spectral encoder and a three-stage fractal progressive growth decoder connected by three layers of cross-stage jumps. The spectral encoder is composed of three cascaded spectral neural network layers, which extracts the global manifold invariants of the incomplete topology in the frequency domain. The three-stage fractal progressive growth decoder dynamically adjusts the splitting factor with curvature as a spatial feedback signal. The adversarial multi-scale discriminator simultaneously evaluates the generated point cloud at both the global topological consistency scale and the local microscopic geometric texture scale.

[0009] Specifically, the spectral graph neural network layer uses the K-nearest neighbor algorithm to construct a point cloud graph structure, performs Chebyshev polynomial approximate spectral convolution on the graph Laplacian matrix in the frequency domain, and implements step-by-step downsampling by sampling the farthest point after each layer. The graph convolution kernel weight matrix adopts a row-level sparse allocation strategy to allocate thread bundles proportionally according to the node degree.

[0010] Specifically, the density adaptive adjustment function is based on the local area point density estimate. Normal vector consistency index and the difference in features between adjacent levels Using the input, calculate the comprehensive adjustment index. According to the adjustment indicators The dynamic given spectrum tensor kernel width parameter within the given interval .

[0011] Among them, the adjustment index With the spectral tensor kernel width parameter In the corresponding relationship, the low threshold of the adjustment index is 0.8, the medium threshold is 2.0, and the high threshold is 5.0, corresponding to the tensor kernel width parameter of the spectral graph. Take 0.05 in sequence. 0.10 0.20 0.40 The boundary values ​​of each interval are determined by orthogonal experiments.

[0012] Specifically, the enhanced point cloud matching framework's three-layer pyramid multi-resolution iterative nearest-point registration involves: Level 1 performing fast coarse registration on point clouds with a retention ratio equal to the coarse registration retention value; Level 2 performing medium-precision registration on point clouds with a medium retention ratio; and Level 3 performing high-precision fine registration on the complete point cloud. The registration result of each layer serves as the initial value for the next layer, and the hybrid distance metric adaptively adjusts the weights of point-to-point distance and point-to-surface distance based on local curvature characteristics.

[0013] In the three-layer pyramid multi-resolution iterative nearest point registration, the retention ratio of point cloud in the first layer is 1%, the retention ratio of point cloud in the second layer is 10%, and the retention ratio of point cloud in the third layer is a complete point cloud. The retention ratio of each layer is determined iteratively by using the registration success rate and convergence speed as evaluation indicators in the registration experiment with different initial pose deviations.

[0014] Specifically, the decoupling algorithm between the physical force equation of the catenary and the spatial variational principle of cable sag deformation involves projecting the cable point cloud along the main span direction onto the vertical plane, constructing a global energy functional, iteratively optimizing the catenary parameters by applying the Euler-Lagrange equations to the functional, decomposing the residual deviation of the point cloud points from the optimal catenary into thermal elongation components and structural deformation components, and outputting the structural deformation residual after deducting the thermal elongation components.

[0015] Wherein, the global energy functional In, weighting coefficient , , The results were determined by least-squares regression analysis on experimental data of known-deformation conductors collected under different wind speed and temperature conditions. The experimental coverage ranged from -20°C to 50°C and from 0°C to 30°C. .

[0016] In the density-based spatial clustering, the neighborhood radius reference value is 2. The minimum sample size reference value is 10. The above parameters were determined by parameter sensitivity experiments on measured data in different vegetation densities and structure distribution scenarios, with the goal of maximizing the crossover ratio between the clustering results and manually labeled hazard areas.

[0017] The training of the curvature flow modulation fractal progressive reconstruction model uses a weighted combination of chamfer distance loss and Earth movement distance loss as the main loss function, with a weight ratio of 7:3. Geometric constraint loss is also introduced, and an adaptive moment estimation optimization algorithm is used. The training dataset has no less than 5,000 sample pairs and is divided into training set, validation set and test set in an 8:1:1 ratio.

[0018] In the training of the curvature flow modulation fractal progressive reconstruction model, the initial learning rate is... The batch size is 16, and the total number of training rounds is 200. The above parameters are determined by a learning rate warm-up experiment and a batch size sensitivity experiment on the validation set. The optimal parameter combination is selected by the comprehensive score of the convergence speed of the chamfer distance on the validation set and the final value.

[0019] A second aspect of the present invention provides a computer-readable storage medium storing program instructions, which, when executed in a computer, are used to perform the above-described method for intelligent identification and early warning of hidden dangers in power transmission lines.

[0020] A third aspect of the present invention provides an intelligent identification and early warning system for potential hazards in power transmission lines, comprising the aforementioned computer-readable storage medium, wherein the system is a computer, the computer-readable storage medium is disposed within the system, and the system is provided with a microprocessor for executing program instructions stored in the computer-readable storage medium.

[0021] This invention constructs a curvature flow modulation fractal progressive reconstruction model and uses a spectral encoder to extract global manifold invariants of the incomplete topology in the frequency domain. This ensures that the overall curvature trend, principal axis direction, and symmetry of the tower point cloud are fully preserved around the missing region. As a result, the completion process does not depend on the local neighborhood features of the missing region, thus avoiding the fundamental defect of spatial domain methods that cause feature breakage due to local missing features.

[0022] This invention further employs a fractal progressive growth decoder, using curvature values ​​as spatial feedback signals to dynamically adjust the splitting factor. Dense branch growth is activated in high-curvature geometric abrupt regions such as crossarm connections and node plates, while redundant splitting is suppressed in low-curvature regions such as the main material. This ensures that the local geometric fidelity and overall distribution consistency of the completed result mutually promote each other. Furthermore, an adversarial multi-scale discriminator constrains the generated point cloud to closely match the real point cloud at the statistical distribution level, guaranteeing that the completed point cloud can be directly used for subsequent registration and deformation analysis.

[0023] In summary, this invention solves the technical problem mentioned in the background art, where partial obstruction of transmission line point cloud data leads to the inability to complete incomplete point cloud data with high fidelity, resulting in misjudgments in subsequent registration and deformation analysis. Attached Figure Description

[0024] Figure 1 This is a flowchart of the method of the present invention.

[0025] Figure 2 The distribution map of characteristic values ​​for registration quality assessment of tower point cloud.

[0026] Figure 3 This is a diagram showing the spatial distance relationship between vegetation hazard clusters and cable point cloud data.

[0027] Figure 4 This is a statistical chart showing the distribution of risk levels across the entire channel. Detailed Implementation

[0028] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below.

[0029] like Figure 1 The diagram shown is a flowchart of a method for intelligent identification and early warning of hidden dangers in power transmission lines provided by the first aspect of the present invention. This method includes the following steps: S01. Collect 3D point cloud data of power transmission channels through a mobile platform equipped with laser scanning equipment, construct a distributed external storage dual-layer decoupled index, divide the channel into geographical data units according to span and tower location, the upper layer adopts a lightweight spatial grid index macro topology, and the lower layer dynamically swaps in and out local point cloud files based on memory mapping technology. S02. Perform a multi-pulse echo full waveform spatiotemporal correlation separation algorithm on the point cloud within the geographic data unit, jointly analyze the laser echo amplitude, pulse half-width and full width at half-maximum and time series energy distribution, construct a single-point physical property scattering matrix, remove secondary scattering artifact points, and then use statistical outlier removal and voxel resampling equalization processing to extract tower point cloud and cable point cloud; S03. The missing tower point cloud and fitting point cloud obtained in step S02 are input into the curvature flow modulation fractal progressive reconstruction model to perform high-fidelity completion of the missing regions. The curvature flow modulation fractal progressive reconstruction model is based on deep learning and outputs a complete tower point cloud. During the completion process, the spectral tensor kernel width parameter is dynamically given by the density adaptive adjustment function based on the point density estimate of the local region, the normal vector consistency index, and the feature difference degree of adjacent levels. S04. Using an enhanced point cloud matching framework based on feature descriptors, the complete tower point cloud obtained in step S03 is registered with the benchmark design point cloud using a three-layer pyramid multi-resolution iterative nearest point registration to obtain an accurate spatial transformation matrix. The registration quality is evaluated based on the eigenvalue analysis of the covariance matrix. S05. Based on the spatial transformation matrix obtained in step S04, extract the tower translation vector and rotation angle. Perform the catenary physical force equation and spatial variation principle cable sag deformation decoupling algorithm on the cable point cloud obtained in step S02, and output the deformation residual as the hazard feature quantitatively. Construct an octree three-dimensional spatial index for the cable point cloud, generate a preliminary screening search sphere point by point along the conductor, and then use a K-dimensional tree to accurately calculate the minimum Euclidean distance from the vegetation point cloud and the structure point cloud to the cable point cloud. Perform density-based spatial clustering on the over-limit points and merge the hazard areas. S06. Establish a multi-factor weighted comprehensive risk assessment model, extract tower deformation parameters, cable spacing parameters, and characteristic parameters of hidden danger areas, introduce voltage level, ambient temperature, and wind speed correction coefficients to dynamically adjust the safe distance threshold, divide the risk level into four levels: low risk, medium risk, high risk, and extremely high risk, construct an interactive three-dimensional visualization scene based on WebGL technology, render dynamic early warning symbols according to the risk level and generate structured early warning reports, and push them to the inspection management system in real time through the application programming interface.

[0030] The principle of the distributed external storage dual-layer decoupled indexing mechanism is as follows: When traditional octrees or K-dimensional trees perform global indexing on billions of point cloud data, memory consumption increases exponentially with the data scale, and multi-level deep traversal introduces severe computational latency. The dual-layer decoupled indexing divides the transmission channel into highly cohesive, loosely coupled independent geographic data units in geographic space. Each geographic data unit corresponds to a complete span or a subset of point clouds between adjacent tower locations. The granularity is jointly determined by the span length and sampling density, with a reference granularity of no more than 512 points per point cloud file corresponding to a single span. The upper-layer lightweight spatial grid stores only the bounding box coordinates and file pointers of each geographic data unit, residing entirely in memory for rapid macro-topological location positioning, reducing query complexity to constant orders of magnitude. The lower layer utilizes memory mapping technology, where the operating system directly maps the disk address space of the point cloud file to the process's virtual address space. When a computational task accesses a geographic data unit, the operating system swaps the corresponding page into physical memory as needed, and automatically swaps it out after computation. Peak physical memory usage is constrained to the product of the number of currently active geographic data units and the size of a single unit's page, fundamentally eliminating the risk of memory overflow. Computational power is decoupled between geographic data units, supporting multi-processor parallel scheduling. Each processor independently holds a memory-mapped handle for its local geographic data unit, avoiding global lock contention. Reference values ​​for geographic data unit segmentation granularity were obtained through iterative experiments on power transmission channel datasets of different sizes: 10, 20, 50, 100, and 200. There are five channel lengths available, with 256 channels per unit. On a memory-intensive server, peak memory usage and single query response time were tested at different granularities, with peak memory usage not exceeding 80% of physical memory and single query response time not exceeding 50 seconds. For the dual constraints, the maximum granularity of the segmentation that satisfies the constraints is taken as the reference value.

[0031] The principle of the multi-pulse echo full-waveform spatiotemporal correlation separation algorithm is as follows: When a laser scanning device irradiates closely adjacent parallel small targets such as split conductors, fittings, and insulator strings, a single laser pulse undergoes multipath scattering on the target surface. Multiple scattered echoes are superimposed on the time axis, forming a synthetic echo with abnormal amplitude and broadened pulse width. Directly analyzing the three-dimensional coordinates will produce false burr point clouds or geometric adhesion between adjacent conductors. The full-waveform spatiotemporal correlation separation algorithm decomposes the above-mentioned aliasing at the physical energy level. The algorithm first performs Gaussian decomposition on the full-waveform digital record, fitting the synthetic echo as the sum of several single-peak Gaussian components. The peak time of each component corresponds to the distance of an independent scattering event. The component amplitude and pulse full width at half maximum (FWHM) jointly characterize the reflectivity and scattering cross-section of the target surface. Thus, an independent single-point physical property scattering matrix is ​​established for each scattering event. The matrix elements include peak amplitude, FWHM, arrival time, and amplitude attenuation rate. Subsequently, a spatiotemporal coherence constraint is introduced: for spatially adjacent scan points, their physical property scattering matrices should satisfy energy continuity in the time series, meaning the difference in arrival time between the scattering components corresponding to the same target on adjacent scan lines should not exceed the theoretical upper limit of time difference determined by the scanning geometry. Components exceeding the upper limit are judged as secondary scattering artifacts and are discarded. Finally, the algorithm reassigns the point cloud to the corresponding physical target based on the time and amplitude of each scattering component, achieving precise boundary separation of fine parallel split wires and eliminating geometric adhesion and false spikes. The upper limit of the number of components used in Gaussian decomposition is determined experimentally: in a laboratory environment, for intervals of 10–80... The parallel steel cable was scanned in full waveform. The number of Gaussian components was gradually increased and the Akaike Information Criterion was used as the model selection criterion. The number of components corresponding to the minimum criterion value was taken as the upper limit reference value. The experimental results showed that the upper limit reference value was 6 components.

[0032] The specific structure of the curvature flow modulation fractal progressive reconstruction model is as follows: the overall architecture of the curvature flow modulation fractal progressive reconstruction model consists of a spectral encoder and a three-stage fractal progressive growth decoder connected by three layers of cross-stage jumps. The spectral encoder consists of three cascaded spectral neural network layers. Each spectral neural network layer first constructs a point cloud structure using the K-nearest neighbor algorithm, with K set to 20. Then, it performs Chebyshev polynomial approximate spectral convolution on the graph Laplacian matrix in the frequency domain, with the polynomial order set to 6, thereby extracting the global manifold invariants of the incomplete topology in the frequency domain, effectively avoiding feature breaks caused by local missing features in the spatial domain. Each spectral neural network layer is followed by sampling the farthest point with a stride of 2 to achieve progressive downsampling, gradually compressing the incomplete point cloud into a global feature vector of dimension 512. The graph convolution kernel weight matrix inside the spectral neural network layer adopts a row-level approach. The sparse allocation strategy allocates CUDA thread bundles proportionally based on the degree of the corresponding node in each row. This ensures that the ratio of thread bundle occupancy for high-degree nodes to low-degree nodes is consistent with the ratio of their degrees, thereby balancing the memory bandwidth usage of each stream processor and eliminating CUDA stream scheduling bubbles caused by sparse and uneven graph structures. The specific thread bundle allocation ratio is determined by benchmark experiments on graph structures with different sparsity. The experiments were conducted on 8 GPUs to test the throughput of a point cloud dataset with a graph node degree distribution covering 5 to 200. The goal was to minimize the standard deviation of the utilization balance of all stream processors to determine the reference value for the thread bundle allocation ratio. Each stage of the three-stage fractal progressive growth decoder includes a feature propagation layer and a fractal growth submodule. The feature propagation layer restores features from low resolution to high resolution through cubic linear interpolation and fuses the skip connection features of the corresponding layer of the encoder. The fractal growth submodule takes the concatenated vector of the parent node features and the global feature vector as conditional input and predicts the 3D offset of the child points generated by the splitting of each parent point through a three-layer fully connected network. After generating the macroscopic skeleton in the first stage, the fractal growth submodule calculates the Gaussian curvature and average curvature of the currently generated surface and uses the curvature values ​​as spatial feedback signals to input to the second and third stage fully connected layers to dynamically adjust the splitting factor. At the crossarm connection where the absolute value of the Gaussian curvature exceeds the threshold, and in the node plate region, dense branches are activated for exponential splitting. In the main material region with flat curvature, splitting is suppressed to prevent redundancy. The bias term of the neuron activation function in the fully connected layer is allocated layer by layer according to the local density distribution statistics of the point cloud generated in the previous stage. The bias term of the corresponding layer is set to a smaller value in the high density region to tighten the activation range, and the bias term of the corresponding layer is set to a larger value in the low density region to widen the activation range. The bias term allocation relationship is determined by ablation experiments on point clouds with different density gradients. The experiments select point cloud samples with density ratios covering 1:1 to 1:20, and the optimal allocation coefficient is selected with the uniformity of activation rate of each layer as the evaluation index.The adversarial multi-scale discriminator evaluates the overall shape distribution at the global topological consistency scale and the surface details at the local micro-geometric texture scale. The Nash equilibrium game between the generator and the discriminator forces the reconstruction result to simultaneously satisfy the macro-structural rationality and the micro-texture authenticity.

[0033] The specific steps for establishing the training dataset of the curvature flow modulation fractal progressive reconstruction model include: selecting qualified complete point cloud samples of towers and fittings from publicly available point cloud datasets and actual collected complete point clouds of transmission channels; randomly simulating occlusion areas for each complete point cloud sample according to the actual scanning occlusion rules; forming input-output training pairs between the simulated occluded point cloud and the corresponding complete point cloud; calculating the Gaussian curvature, average curvature, and normal vector label values ​​for each point cloud sample; constructing a training dataset of no less than 5000 sample pairs; and dividing the training set, validation set, and test set in an 8:1:1 ratio.

[0034] The specific steps of training the curvature flow modulation fractal progressive reconstruction model include: using a weighted combination of chamfer distance loss and Earth movement distance loss as the main loss function, with a weight ratio of 7:3. The weight ratio is determined by conducting a grid search experiment on the validation set to evaluate the two loss weights, with a search step size of 0.1, and selecting the optimal weight with the minimum chamfer distance on the validation set as the objective; jointly introducing geometric constraint loss to penalize the surface roughness of the generated point cloud; and employing an adaptive moment estimation optimization algorithm with an initial learning rate set to... The batch size was set to 16, and the total number of training epochs was 200. The initial learning rate and batch size reference values ​​were determined jointly by a learning rate warm-up experiment and a batch size sensitivity experiment on the validation set. The warm-up experiment... to Within the range, samples were taken at logarithmic intervals. Batch size sensitivity experiments were conducted one by one at 8, 16, 32, and 64 to verify the optimal parameter combination based on the convergence speed of the set chamfer distance and the comprehensive score of the final value.

[0035] The curvature flow modulation fractal progressive reconstruction model extracts global manifold invariants in the frequency domain through a spectral encoder, ensuring that the structural information of the incomplete topology is fully preserved around the missing regions, fundamentally avoiding the truncation effect of local spatial domain defects on feature extraction. The fractal progressive growth decoder uses curvature flow as the modulation signal, refining the point cloud completion process layer by layer from the macroscopic skeleton to the microscopic details. It grows densely in geometrically abrupt regions such as crossarms and node plates with high curvature, and sparsely in the main material regions with low curvature, so that the local geometric fidelity of the completion result and the consistency of the overall distribution mutually promote each other. The introduction of an adversarial multi-scale discriminator ensures that the generated point cloud maintains a high degree of consistency with the real point cloud at the statistical distribution level, ensuring that the completion result can be directly used for subsequent registration and deformation analysis, fundamentally eliminating the risk of deformation misjudgment caused by point cloud defects.

[0036] The principle of the density adaptive adjustment function is as follows: based on the estimated point density values ​​of the local region... (unit: ), normal vector consistency index (Dimensionless, ranging from 0 to 1) and the difference in features between adjacent levels (Dimensionless, ranging from 0 to 1) is used as input to calculate the comprehensive regulation index. ;when At that time, the kernel width parameter of the spectral tensor Pick This corresponds to a low-curvature, flat region; when hour, Pick This corresponds to a region of moderate geometric complexity; when hour, Pick , corresponding to high curvature regions; when hour, Pick This corresponds to a densely connected region of node plates with extremely high geometric complexity. The boundary values ​​of 0.8, 2.0, and 5.0 for each interval were obtained through piecewise experiments on point cloud samples with different combinations of density and curvature: [Constructing a covering...] The range is indivual , The range is 0.1 to 0.9. For an orthogonal experimental sample set ranging from 0.05 to 0.95, calculations were performed for each parameter combination. The corresponding geometric complexity level is manually labeled, and the segment boundary with the highest classification accuracy is used as the reference value.

[0037] The principle of the decoupling algorithm for cable sag deformation based on the physical force equation of the catenary and the spatial variational principle is as follows: Under the influence of its own weight, wind load, and temperature, the natural sag curve of an overhead conductor satisfies the catenary equation. ; parameters Characterized by the ratio of horizontal tension of the cable to its weight per unit length. Characterizes the horizontal position of the lowest point of the conductor. The algorithm characterizes the elevation of the lowest point of the conductor. First, it projects the cable point cloud along the main span direction onto the plumb plane and estimates the plumb plane normal vector using the least squares method. Within the plumb plane, a global energy functional is constructed, using the sum of squared Euclidean distances between the actual point cloud data and the theoretical catenary as the external energy functional, and the integrals of conductor bending stiffness and tension work as the internal canonical functional. ;in For point cloud points To the catenary curve Euclidean distance, For the curvature of the conductor, For tension, , , These are weighting coefficients. Let be the arc length parameter. Solve the Euler-Lagrange equations for the functional, iteratively optimizing the parameters in the spacetime dimension. , , With weighting coefficients, after convergence, the residual deviation of the point cloud points to the optimal catenary is decomposed into a thermal elongation component along the tangent direction of the conductor and a structural deformation component perpendicular to the tangent direction. The thermal elongation component is estimated and subtracted from the current ambient temperature and the conductor's thermal expansion coefficient. The remaining structural deformation residual is quantitatively output as a hazard characteristic. Weighting coefficients , , The reference values ​​were determined by least-squares regression analysis on experimental data of known-deformation conductors collected under different wind speed and temperature conditions. The experimental coverage ranged from -20 to 50°C in temperature and from 0 to 30 km / h in wind speed. The deformation annotation accuracy is at the millimeter level.

[0038] The proposed catenary physical force equation and spatial variational principle cable sag deformation decoupling algorithm introduces physical mechanism constraints into point cloud deformation analysis, enabling the algorithm to distinguish between normal physical deformation caused by temperature and wind load and abnormal structural deformation caused by strand breakage, icing, and external forces, thus avoiding misjudging normal sag changes as line faults. The variational framework unifies global energy optimization constraints and local point cloud data fitting within the same mathematical system, ensuring that parameter estimation remains robust even with high data noise, and providing highly reliable quantitative deformation feature inputs for subsequent multi-factor risk assessment.

[0039] The enhanced point cloud matching framework employs a three-layer pyramid multi-resolution iterative nearest-point registration. Layer 1 performs rapid coarse registration while retaining 1% of the point cloud; Layer 2 performs medium-precision registration while retaining 10% of the point cloud; and Layer 3 performs high-precision fine registration on the complete point cloud. The registration result of each layer serves as the initial value for the next layer, gradually converging to the global optimum. The hybrid distance metric adaptively adjusts the weights of point-to-point distance and point-to-surface distance based on local curvature characteristics, assigning higher weight to point-to-point distance in high-curvature regions and higher weight to point-to-surface distance in low-curvature regions. The reference values ​​for the point cloud retention ratio of each layer are iteratively determined using registration success rate and convergence speed as evaluation indicators in registration experiments with different initial pose deviations. The initial pose deviation coverage translation ranges from 0 to 500. The rotation angle is in the range of 0 to 30°.

[0040] In the density-based spatial clustering, the reference values ​​for the neighborhood radius and the minimum sample number are determined by parameter sensitivity experiments on measured data from different vegetation densities and structure distribution scenarios: the neighborhood radius is between 1 and 5. Within the range of 0.5 Step-size search, with a minimum sample size of 5-30, uses a step size of 5 to search. The optimal parameter combination is selected with the highest intersection-union ratio (IUU) between the clustering results and the manually labeled hazard areas. The reference value for the neighborhood radius is 2. The minimum sample size reference value is 10.

[0041] The spectral graph neural network refers to a graph neural network that performs convolution operations on the frequency domain of the Laplacian matrix of a graph. By approximating the spectral convolution with Chebyshev polynomials, the convolution kernel has global manifold awareness capabilities and is suitable for three-dimensional point cloud maps with irregular topological structures.

[0042] The global manifold invariant refers to the low-frequency characteristics that remain stable in the frequency domain when the topology of the point cloud is locally missing or disturbed, including macroscopic geometric properties such as the overall bending trend of the point cloud, principal axis direction, and symmetry.

[0043] The geographic data unit refers to a highly cohesive, low-coupling subset of point clouds that is divided in geographic space according to the span and tower location. Each geographic data unit is stored independently as a disk file, serving as the basic scheduling unit of the distributed external storage index.

[0044] The single-point scattering matrix refers to a matrix structure composed of four physical quantities: peak amplitude, pulse full width at half maximum (FWHM), arrival time, and amplitude attenuation rate of a single laser scattering event. It is used to characterize the scattering characteristics of the target surface from the physical energy layer and to support the separation and judgment of multipath echoes.

[0045] Wherein, the normal vector consistency index The calculation method is as follows: Within the neighborhood sphere centered on the point to be evaluated, calculate the mean cosine similarity between the normal vectors of all neighboring points and the normal vector of the center point. The mean cosine similarity is... The closer the value is to 1, the flatter the local surface is.

[0046] Among them, the feature difference between adjacent levels The calculation method is as follows: Calculate the normalized Euclidean distance between the output feature vectors of two adjacent layers of the spectral encoder. The normalization method is to divide by the square root of the feature vector dimension, and the resulting value is mapped to the range of 0 to 1. The larger the value, the more drastic the change in geometric structure between adjacent levels.

[0047] The vertical plane refers to the plane spanned by the main span direction vector of the conductor and the vertical direction vector. The sag deformation of the conductor under gravity mainly occurs within the vertical plane. Projecting the point cloud onto the vertical plane can effectively separate the horizontal displacement caused by wind deflection from the vertical displacement caused by sag.

[0048] The specific implementation of step S01 is as follows: A drone or vehicle-mounted platform equipped with a laser scanning device is used to scan the power transmission channel, acquiring raw point cloud data containing three-dimensional coordinates, echo intensity, and full waveform information. To efficiently manage the channel point cloud at the billion-level scale, a distributed external storage dual-layer decoupled index is constructed. The upper-layer lightweight spatial grid divides the channel into several geographic data units according to span and tower location. Each geographic data unit covers a complete span, with a reference granularity where the point cloud file size corresponding to a single span does not exceed 512. The upper layer stores only the bounding box coordinates and disk file pointers of each geographic data unit, residing entirely in memory and supporting constant-order macro-topology queries. The lower layer, based on memory mapping technology, directly maps the disk address space of the point cloud file to the process's virtual address space. When a computation task accesses a geographic data unit, the corresponding page is swapped in as needed and automatically swapped out after computation. Peak physical memory usage is constrained to the product of the number of active geographic data units and the size of a single unit's page. Computational power is decoupled between geographic data units; each processor independently holds a memory-mapped handle to its local geographic data unit, supporting multi-processor parallel scheduling, avoiding global lock contention, and fundamentally eliminating the risk of memory overflow.

[0049] The specific implementation of step S02 is as follows: A multi-pulse echo full-waveform spatiotemporal correlation separation algorithm is executed on the original full-waveform point cloud data within each geographic data unit. The purpose is to accurately separate the aliased echoes of closely adjacent parallel small targets such as split conductors, hardware, and insulator strings. The algorithm first performs Gaussian decomposition on the full-waveform digitized record, fitting the synthesized echo as the sum of several single-peak Gaussian components. The upper limit of the number of components is referenced as 6, and the peak time of each component corresponds to the distance of an independent scattering event. A single-point physical property scattering matrix is ​​established for each scattering event based on four physical quantities: peak amplitude, pulse full width at half maximum (FWHM), arrival time, and amplitude attenuation rate. The matrix elements comprehensively characterize the scattering properties of the target surface. Subsequently, a spatiotemporal coherence constraint is introduced. For the physical property scattering matrices of spatially adjacent scan points, the difference in arrival time between the corresponding scattering components of the same target is checked to see if it exceeds the theoretical upper limit of time difference determined by the scanning geometry. Components exceeding the upper limit are judged as secondary scattering artifacts and discarded. Finally, based on the time and amplitude of each scattering component, the point cloud is reassigned to the corresponding physical target. Then, statistical outlier removal and voxel resampling equalization are used to extract the tower point cloud and cable point cloud.

[0050] The specific implementation of step S03 is as follows: The incomplete tower point cloud and fitting point cloud obtained in step S02 are input into the curvature flow modulation fractal progressive reconstruction model for high-fidelity completion. The overall model architecture consists of a spectral encoder and a three-stage fractal progressive growth decoder connected by three layers of cross-stage jumps. The spectral encoder consists of three cascaded spectral neural network layers. Each layer constructs the point cloud structure using the K-nearest neighbor algorithm (K value is 20). Chebyshev polynomial approximate spectral convolution (polynomial order is 6) is performed on the graph Laplacian matrix in the frequency domain to extract the global manifold invariants of the incomplete topology. Each layer is followed by sampling the farthest point with a stride of 2, gradually compressing the incomplete point cloud into a 512-dimensional global feature vector. The graph convolution kernel weight matrix adopts a row-level sparse allocation strategy, allocating thread bundles proportionally according to node degree to balance the memory bandwidth usage of the streaming processor. Each stage of the three-stage fractal progressive growth decoder includes a feature propagation layer and a fractal growth submodule. The fractal growth submodule takes the concatenated vector of the parent node features and the global feature vector as conditional input and predicts the 3D offset of the sub-points through a three-layer fully connected network. After generating the macroscopic skeleton in the first stage, the Gaussian curvature and average curvature of the currently generated surface are calculated. The curvature values ​​are used as spatial feedback signals to input the second and third stages. Dense branches are activated at crossarm connections and node plate regions where the absolute value of the Gaussian curvature exceeds the threshold, resulting in exponential splitting. Splitting is suppressed in the main material regions with flat curvature. The bias term of the fully connected layer is allocated layer by layer according to the local density distribution statistics of the point cloud generated in the previous stage. Smaller bias values ​​are used for high-density regions, and larger bias values ​​are used for low-density regions. The adversarial multi-scale discriminator simultaneously constrains the generated point cloud at the global topological consistency scale and the local micro-geometric texture scale, forcing the completion result to simultaneously satisfy the rationality of the macroscopic structure and the authenticity of the micro-texture. The spectral tensor kernel width parameter is dynamically given by the density adaptive adjustment function, which uses the local region point density estimate. Normal vector consistency index and the difference in features between adjacent levels Using the input, calculate the comprehensive adjustment index. ,when Take 0.05 at a time ,when Take 0.10 at a time. ,when Take 0.20 at a time ,when Take 0.40 at the time. .

[0051] The specific implementation of step S04 is as follows: An enhanced point cloud matching framework based on feature descriptors is used to perform three-layer pyramid multi-resolution iterative nearest-point registration with the complete tower point cloud obtained in step S03 and the benchmark design point cloud. Level 1 performs fast coarse registration while retaining 1% of the point cloud; Level 2 performs medium-precision registration while retaining 10% of the point cloud; and Level 3 performs high-precision fine registration on the complete point cloud. The registration result of each layer serves as the initial value for the next layer, gradually converging to the global optimum. The hybrid distance metric adaptively adjusts the weights of point-to-point distance and point-to-area distance based on local curvature characteristics, assigning higher weight to point-to-point distance in high-curvature regions and higher weight to point-to-area distance in low-curvature regions. After registration, the registration quality is evaluated based on covariance matrix eigenvalue analysis, and an accurate spatial transformation matrix is ​​output.

[0052] The specific implementation of step S05 is as follows: Based on the spatial transformation matrix obtained in step S04, the tower translation vector and rotation angle are extracted. The catenary physical force equation and the spatial variational principle cable sag deformation decoupling algorithm are then applied to the cable point cloud obtained in step S02. The algorithm first projects the cable point cloud along the main span direction onto the vertical plane. The least squares method is used to estimate the vertical plane normal vector. Within the vertical plane, the sum of squared Euclidean distances between the actual point cloud data and the theoretical catenary is used as the external energy functional, and the integral of conductor bending stiffness and tension work is used as the internal regular functional, thus constructing a global energy functional. Find the Euler-Lagrange equation for the functional and iteratively optimize the catenary parameters. , , After convergence, the residual deviation is decomposed into thermal elongation and structural deformation components using weighted coefficients. The structural deformation residual is output as a hazard characteristic after subtracting the thermal elongation component. An octree 3D spatial index is constructed for the cable point cloud. Initial search spheres are generated point-by-point along the conductor. Then, a K-dimensional tree is used to accurately calculate the minimum Euclidean distance from the vegetation point cloud and structure point cloud to the cable point cloud. Density-based spatial clustering is performed on the out-of-limit points, with a neighborhood radius reference value of 2. The minimum sample size reference value is 10, and the clustering results are merged into the hidden danger area.

[0053] The specific implementation of step S06 is as follows: A multi-factor weighted comprehensive risk assessment model is established. Using tower deformation parameters, cable spacing parameters, and characteristic parameters of the hazard area as inputs, voltage level, ambient temperature, and wind speed correction coefficients are introduced to dynamically adjust the safe distance threshold, classifying the risk level into four levels: low risk, medium risk, high risk, and extremely high risk. An interactive 3D visualization scene is constructed based on WebGL technology. Dynamic warning symbols are rendered according to the risk level, generating a structured warning report containing the risk level, hazard type, hazard location, and suggested measures. This report is pushed to the inspection management system in real time via an application programming interface, achieving a fully automated closed-loop hazard warning process.

[0054] It should be noted that the key technologies of this invention include: a spectral encoder extracts global manifold invariants in the frequency domain, ensuring that the macroscopic geometric properties of the incomplete topology are fully preserved around the missing region, overcoming the defect that local features in the spatial domain cannot transmit topological information due to the breakage of the missing region, and providing a stable global prior for high-fidelity point cloud completion; a fractal progressive growth decoder uses curvature flow as the modulation signal, enabling adaptive matching between the point cloud density distribution and the target geometric complexity, growing densely in high curvature regions and sparsely in low curvature regions, so that the local geometric fidelity of the completion result and the consistency of the overall distribution mutually promote each other; and the joint constraint of the catenary physical force equation and the spatial variational principle decouples normal sag changes and abnormal structural deformations at the mathematical level, giving deformation feature extraction a physical mechanism support. The synergistic effect of these three key technologies enables the entire method to form a high-precision closed-loop technology chain from data acquisition, point cloud completion, deformation analysis to risk warning, significantly reducing the rate of missed and false alarms caused by point cloud missing and sag misjudgment.

[0055] It should be noted that in overhead transmission lines, large areas of point cloud loss are easily generated during laser scanning due to shading at crossarm attachment points, densely connected areas of node plates, and insulator connection points. These missing areas are often precisely where structural deformation is most sensitive and requires the most accurate modeling. Under these conditions, accurately completing the point cloud in the missing areas while ensuring that the completion result does not introduce additional errors in subsequent registration and deformation analysis is a highly challenging technical problem. The reason for this technical problem is that the local neighborhood features of the missing area are completely broken. The spatial domain local feature extraction operator can only perceive a limited neighborhood near the missing boundary and cannot obtain global geometric priors such as the overall bending trend and symmetry of the tower. This leads to defects such as topological breaks, surface wrinkles, and uneven point cloud distribution in the completion result in areas of high curvature geometric abrupt change. The registration error between the completed point cloud and the baseline design point cloud increases significantly, thus misjudging normal structures as potential deformation hazards. The common solution to the aforementioned technical problems is to employ surface reconstruction algorithms based on Poisson reconstruction or moving least squares, which fill in the missing regions by interpolating or extrapolating neighborhood points around the missing boundaries. However, Poisson reconstruction relies on continuously differentiable implicit functions, and the boundary conditions of these implicit functions are unreliable when there are large areas of missing data, resulting in smoothing distortion in the reconstruction results. Moving least squares only utilizes a limited number of neighborhood points near the missing boundaries and cannot perceive the global topological constraints at geometrically abrupt locations, leading to a lack of overall consistency in the completion results and failing to meet the requirements for subsequent high-precision registration. This invention effectively solves this technical problem by performing Chebyshev polynomial approximate spectral convolution on the frequency domain of the graph Laplacian matrix using a spectral encoder. The low-frequency components in the frequency domain correspond to the global manifold structure of the point cloud, which remains stable even when there are local missing data. This allows the encoder to extract complete global manifold invariants from the incomplete point cloud, including the overall bending trend of the tower, the principal axis direction, and symmetry, providing a reliable global prior for the decoding stage. This ensures that the completion process does not depend on the local neighborhood features of the missing region. The fractal progressive growth decoder uses curvature flow as the modulation signal to activate dense branch growth in regions of high curvature geometric abrupt changes and suppress redundant splitting in regions of low curvature, ensuring high geometric fidelity in the completion results at key locations such as crossarms and node plates. The introduction of an adversarial multi-scale discriminator ensures that the generated point cloud closely matches the real point cloud in terms of statistical distribution, guaranteeing that the completed point cloud can be directly used for subsequent registration and deformation analysis, thereby eliminating the risk of deformation misjudgment caused by missing point clouds.

[0056] A second aspect of the present invention provides a computer-readable storage medium storing program instructions, which, when executed in a computer, are used to perform the above-described method for intelligent identification and early warning of hidden dangers in power transmission lines.

[0057] A third aspect of the present invention provides an intelligent identification and early warning system for hidden dangers in power transmission lines, comprising the aforementioned computer-readable storage medium. The system can be any one of a computer, a server, or a microcontroller. The computer-readable storage medium is disposed within the system, and the system is provided with a microprocessor that executes the program instructions stored in the computer-readable storage medium.

[0058] Specifically, the principle of this invention is: The fundamental reason why this invention can solve the above-mentioned technical problems is that the spectral encoder performs Chebyshev polynomial approximate spectral convolution on the frequency domain of the Graph Laplacian matrix. The low-frequency components in the frequency domain correspond to the global manifold structure of the point cloud, which remains stable even when local regions are missing and will not be lost due to spatial domain neighborhood breaks. This property enables the encoder to extract complete global manifold invariants from the incomplete point cloud, including the overall bending trend of the tower, principal axis direction, and symmetry, providing a reliable global prior for the subsequent decoding stage.

[0059] The fractal progressive growth decoder decomposes the completion process into three stages of progressive refinement, from macroscopic skeleton to microscopic details. Each stage uses the Gaussian curvature and mean curvature of the surface generated in the previous stage as spatial feedback to dynamically adjust the splitting factor of the current stage. High curvature regions correspond to geometrically abrupt locations such as crossarm connections and node plates, requiring dense point clouds to accurately represent the local shape, thus activating exponential splitting. Low curvature regions correspond to flat locations such as the main material, where sparse splitting can meet the accuracy requirements while avoiding the introduction of redundant points. This curvature modulation mechanism adaptively matches the point cloud density distribution with the target geometric complexity, thereby achieving a balance between global fidelity and local detail.

[0060] The density adaptive adjustment function calculates the comprehensive adjustment index based on the local point density estimate, the normal vector consistency index, and the feature difference between adjacent layers. It dynamically gives the spectral tensor kernel width parameter, so that the spectral convolution kernel has different sensing ranges in regions with different geometric complexities, further enhancing the encoder's adaptability to non-uniform point clouds.

[0061] An adversarial multi-scale discriminator simultaneously evaluates the generated point cloud at both the global topological consistency scale and the local microscopic geometric texture scale. Through a Nash equilibrium game between the generator and the discriminator, the completed result is forced to simultaneously satisfy macroscopic structural rationality and microscopic texture realism, ensuring that the completed point cloud highly matches the real point cloud at the statistical distribution level. The completed tower point cloud is then aligned with the benchmark design point cloud through three-layer pyramid multi-resolution iterative nearest-point registration, significantly reducing the registration error. This allows subsequent tower deformation parameter extraction and cable sag deformation decoupling analysis to be based on a high-precision spatial transformation matrix, avoiding deformation misjudgments caused by missing point clouds.

[0062] The following provides a specific embodiment 1 of the present invention, and the specific implementation of each step in this embodiment 1 is described in detail below.

[0063] The specific implementation of step S01 involves collecting 3D point cloud data of the power transmission channel via a mobile platform equipped with a laser scanning device, and then constructing a distributed external storage dual-layer decoupled index. The upper-layer lightweight spatial grid stores only the bounding box coordinates and file pointers of each geographic data unit, residing entirely in memory, reducing query complexity to constant levels. The lower layer, based on memory mapping technology, directly maps the disk address space of the point cloud file to the process's virtual address space, swapping in and out as needed. Peak physical memory usage is constrained to the product of the number of currently active geographic data units and the page size of a single unit. Reference values ​​for geographic data unit segmentation granularity are set at 10, 20, 50, 100, and 200. The results were obtained through iterative experiments on datasets with five different channel lengths, with peak memory usage not exceeding 80% of physical memory and single query response time not exceeding 50 seconds. For the dual constraints, the largest segmentation granularity that satisfies the constraints is selected. The reference granularity is that the point cloud file size corresponding to a single segment should not exceed 512. .

[0064] The specific implementation of step S02 is to perform a multi-pulse echo full-waveform spatiotemporal correlation separation algorithm on the point cloud within the geographic data unit. First, Gaussian decomposition is performed on the full-waveform digitized record to synthesize the echo. The fit is the sum of several unimodal Gaussian components, expressed by the following formula: ; In the formula, For time, The upper limit of the number of Gaussian components. For the first The peak amplitude of each component For the first The peak arrival time of each component, For the first The pulse width parameter of each component, This represents the fitting error term. The value of is determined by the Akaike Information Criterion, and the formula is expressed as follows: ; In the formula, The number of components is The Akaike Information Criterion Value at that time For the corresponding maximum likelihood estimate, take The minimum corresponding The value is used as the upper limit for the number of components, and the upper limit reference value was determined to be 6 through experiments. A single-point scattering matrix is ​​established for each scattering event. The formula is expressed as follows: ; In the formula, For the first The amplitude attenuation rate of each component, through the analysis of... The first-order difference estimate of time is obtained. Subsequently, a spatiotemporal coherence constraint is introduced, requiring the time difference between the arrival times of the scattering components of adjacent scan points to satisfy: ; In the formula, and Adjacent scan points and Corresponding to the The arrival time of the component, The theoretical upper limit of the time difference is determined by the scanning geometry, and is based on the spatial distance between adjacent scanning points. with the speed of light The joint estimation is expressed by the following formula: ; In the formula, Adjacent scan points and Spatial distance between them At the speed of light, This is the system time jitter compensation amount, with an empirical value of [value missing]. Components exceeding the upper limit are identified as secondary scattering artifacts and removed. Subsequently, statistical outlier removal and voxel resampling equalization are performed to extract tower and cable point clouds.

[0065] The specific implementation of step S03 is to perform high-fidelity completion on the curvature flow modulation fractal progressive reconstruction model of the incomplete tower point cloud input. The spectral encoder consists of three cascaded spectral neural network layers, each layer using the K-nearest neighbor algorithm (KNN). Construct a point cloud graph structure and perform a graph Laplacian matrix in the frequency domain. Perform Chebyshev polynomial approximation spectral convolution, as expressed in the following formula: ; In the formula, For the first Layer node feature matrix, Let be the order of the polynomial (take 6). For the first Layer Learnable weights of order, For the normalized graph Laplace matrix, for The Chebyshev polynomials For the first Layer bias term, From the recursive relation Calculation, initial conditions are , , The identity matrix is ​​used. Each layer is followed by sampling of the farthest point with a step size of 2, progressively compressing the residual feature cloud into a global feature vector of dimension 512. The fractal growth submodule is based on the characteristics of the parent node. With global feature vectors The concatenated vector is used as the conditional input, and a three-layer fully connected network is used to predict the 3D offset of the child points of each parent point. The formula is expressed as follows: ; In the formula, Indicates the first Fully connected layer This represents a vector concatenation operation. For the first Feature vectors of each parent point. Comprehensive adjustment index. Calculated using the density adaptive adjustment function, the formula is expressed as follows: ; In the formula, Point density estimate for a local area (unit: ), The normal vector consistency index. The difference in features between adjacent levels. This is a comprehensive adjustment index (dimensionless). When... Temporal spectrum tensor kernel width parameter Take 0.05 ,when hour Take 0.10 ,when hour Take 0.20 ,when hour Take 0.40 , Compared with the Gaussian pulse width parameter in step S02 The meanings are different. Normal vector consistency index It is obtained by calculating the mean cosine similarity between the normal vectors of all neighboring points within the neighborhood sphere and the normal vector of the center point, as expressed in the following formula: ; In the formula, For the number of neighboring points, and These are the unit normal vectors of the center point and the neighboring points, respectively. Center point The neighborhood point set. Distinction between adjacent hierarchical features. The formula for calculating using normalized Euclidean distance is as follows: ; In the formula, The Euclidean distance is the output feature vector of two adjacent layers. The dimension of the feature vector. Because... and For real-valued eigenvectors of the same dimension, and Same dimensions This is a dimensionless quantity. The training loss function is a weighted combination of the chamfer distance loss and the Earth movement distance loss, expressed by the following formula: ; In the formula, For chamfer distance loss, Loss due to distance traveled on Earth The geometric constraint loss (a weighted sum of normal vector consistency constraints and point spacing regularity constraints, used to penalize the surface roughness of the generated point cloud). The geometric constraint weight coefficient has an empirical value of 0.1. Weights of 0.7 and 0.3 were determined through grid search experiments (step size 0.1) on the validation set. , , , All dimensions are .

[0066] The specific implementation of step S04 is as follows: An enhanced point cloud matching framework is used to perform a three-layer pyramid multi-resolution iterative nearest-point registration. Layer one performs coarse registration on 1% of the point cloud, layer two performs medium-precision registration on 10% of the point cloud, and layer three performs high-precision fine registration on the complete point cloud. The registration result of each layer is used as the initial value for the next layer. The hybrid distance metric adaptively adjusts the weights of point-to-point distance and point-to-area distance based on local curvature, as expressed in the following formula: ; In the formula, and For the corresponding point pair, 3D coordinate vectors For local Gaussian curvature, For point-to-point distance weighting, Weighted by the distance between points and surfaces. For point The unit normal vector at that location, Dimensions are . and The calculation formula is expressed as follows: ; ; In the formula, The normalized curvature reference value is 0.1, with an empirical value of 0.1. , and All are dimensionless quantities, satisfying Registration quality is determined by the covariance matrix. The eigenvalue analysis and evaluation are expressed by the following formula: ; In the formula, The number of point pairs participating in the registration. For the first The residual vector of each corresponding point For the residual mean vector, Eigenvalue decomposition is performed; when the ratio of the largest to the smallest eigenvalue approaches 1, the registration quality is considered good. The reference values ​​for the point cloud retention ratio at each layer range from a translation amount of 0 to 500. Registration experiments were conducted within the rotation angle range of 0 to 30°.

[0067] The specific implementation of step S05 is as follows: After extracting the tower translation vector and rotation angle based on the spatial transformation matrix, the catenary physical force equation and the spatial variational principle cable sag deformation decoupling algorithm are applied to the cable point cloud. After projecting the point cloud onto the vertical plane, a global energy functional is constructed, expressed by the following formula: ; In the formula, For global energy functional, For point cloud points To the catenary curve Euclidean distance, For the curvature of the conductor, For conductor tension, For arc length parameters, Weighting coefficients for the data fitting term (dimensions: 1) ), Weighting coefficients for the bending stiffness term (dimensions: 1) ), Weighting coefficients for the tension term (dimensions: 1) The dimensions of the three integrals are all... , , , Coverage temperature -20 to 50°C, wind speed 0 to 30 The measured data were determined by least squares regression. Dimensions are The equation for the catenary is expressed as follows: ; In the formula, Let be the vertical coordinates of the traverse in the plumb plane. Let be the horizontal coordinates of the traverse in the plumb plane. This is the ratio of the horizontal tension of the cable to its weight per unit length. The lowest point of the conductor is at a horizontal position. Let the elevation be the lowest point of the traverse. Solve the Euler-Lagrange equations using the functional and iteratively optimize the parameters. , , and , , After convergence, the residual deviation is decomposed into thermal elongation and structural deformation components. Thermal elongation component The estimation formula is expressed as follows: ; In the formula, The coefficient of thermal expansion of the conductor. The original length of the conductor at the reference temperature. The current ambient temperature. This is a reference temperature (the empirical value is 20℃). This is for thermal elongation. Subtract... The remaining structural deformation residuals are then quantitatively output as hazard characteristics. Subsequently, an octree 3D spatial index for the cable point cloud is constructed, and initial screening search spheres are generated point-by-point. Then, a K-dimensional tree is used to accurately calculate the minimum Euclidean distance from the vegetation and structure point clouds to the cable point cloud. Density-based spatial clustering is performed on the out-of-limit points, with a neighborhood radius reference value of 2. The minimum sample size reference value is 10.

[0068] The specific implementation of step S06 is to establish a multi-factor weighted comprehensive risk assessment model, introduce voltage level, ambient temperature and wind speed correction coefficients to dynamically adjust the safe distance threshold, divide the risk level into four levels: low risk, medium risk, high risk and extremely high risk, construct an interactive three-dimensional visualization scene based on WebGL technology, render dynamic early warning symbols according to the risk level and generate a structured early warning report, and push it to the inspection management system in real time through the application programming interface.

[0069] To better understand and implement this invention, a specific application scenario of this invention is provided in Embodiment 2 below: This embodiment applies to a certain 220 The power transmission line corridor is equipped with intelligent hazard identification and early warning systems, with a total corridor length of approximately 85 kilometers. It contains a total of 168 points, and the scanning data was collected by a drone platform equipped with laser scanning equipment. The total amount of raw point cloud data is approximately 420. The average size of the point cloud file corresponding to a single distance is approximately 380. This satisfies the granularity reference constraint of the distributed external storage dual-layer decoupled index.

[0070] In step S01, the channel point cloud is divided into 168 geographic data units according to the span and tower location. Each geographic data unit is stored independently as a disk file. The upper-layer lightweight spatial grid stores 168 sets of spatial boundary bounding box coordinates and file pointers, all residing in memory, with memory usage less than 1 byte. The lower layer uses memory mapping technology for on-demand swapping in and out, thus limiting the server's peak physical memory usage to within the range of active processing units, achieving support for 420 Efficient management of raw data.

[0071] In step S02, a multi-pulse echo full-waveform spatiotemporal correlation separation algorithm is executed for each geographic data unit. Taking a typical span as an example, this span contains four-split conductors with a conductor spacing of approximately 400 mm. During scanning, a single laser pulse simultaneously illuminates adjacent split conductors, generating aliased composite echoes. The algorithm performs Gaussian decomposition on the full waveform digital record, with an upper limit of 6 components. After removing secondary scattering artifacts through spatiotemporal coherence constraints, the point clouds of the four split conductors are accurately separated, eliminating geometric adhesion. After statistical outlier removal and voxel resampling equalization, approximately 1.82 million tower point clouds and approximately 960,000 cable point clouds are extracted within this span.

[0072] In step S03, the extracted incomplete tower point cloud is supplemented by a curvature flow modulation fractal progressive reconstruction model. Taking a certain straight tower as an example, due to flight altitude limitations, there is significant occlusion and missing areas in the densely connected region between the crossarm root and the node plate, with the missing area accounting for approximately 23% of the total tower point cloud area. After the model spectral encoder extracts global manifold invariants in the frequency domain, the three-stage fractal progressive growth decoder uses Gaussian curvature and mean curvature as spatial feedback to activate dense branch growth at the crossarm connection and node plate regions, and suppress redundant splitting in the main material region, ultimately outputting a complete tower point cloud. The quality comparison of the tower point cloud before and after supplementation is shown in Table 1.

[0073] Table 1. Comparison of Pole Point Cloud Computing Quality Before and After Completion

[0074] As shown in Table 1, the point density in the crossarm and node plate regions significantly increased after completion, the average normal vector consistency index increased from 0.61 to 0.93, and the chamfer distance from the baseline model increased from 48.3. Reduced to 6.7 This indicates that the completion effect meets the requirements for subsequent registration.

[0075] In step S04, an enhanced point cloud matching framework is used to perform three-level pyramid multi-resolution iterative nearest-point registration between the complete tower point cloud and the benchmark design point cloud. Level 1 performs coarse registration on 1% of the point cloud (approximately 18,000 points), Level 2 performs medium-precision registration on 10% of the point cloud (approximately 180,000 points), and Level 3 performs high-precision fine registration on the complete point cloud, ultimately obtaining an accurate spatial transformation matrix. Based on covariance matrix eigenvalue analysis, the registration quality assessment results for this tower are as follows: Figure 2 As shown, the eigenvalue ratios in the three translation directions are uniform, indicating that the registration result converges to the global optimum. Based on the spatial transformation matrix, the translation vector of the tower is extracted as (3.2). 1.8 4.5 The rotation angle was 0.12°, which is within the normal range, indicating that the tower has no obvious structural deformation.

[0076] In step S05, the catenary physical force equation and the spatial variational principle cable sag deformation decoupling algorithm are applied to the cable point cloud. Taking a certain span as an example, the horizontal span of the conductor is 386. The ambient temperature during the measurement was 32℃, and the average wind speed was 8. The algorithm projects the cable point cloud onto the vertical plane, constructs a global energy functional, iteratively optimizes the catenary parameters, extracts and subtracts the thermal elongation component after convergence, and the statistical results of the remaining structural deformation residuals are shown in Table 2.

[0077] Table 2. Decoupling Results of Cable Sag Deformation

[0078] As shown in Table 2, the total sag deviation is 187. In the middle, 143 Thermal elongation caused by temperature is a normal physical deformation; the remaining 44 Structural deformation residuals were quantitatively extracted as hazard features and input into subsequent risk assessments. Subsequently, an octree 3D spatial index of the cable point cloud was constructed. The K-dimensional tree search results showed that within this span, there was a minimum Euclidean distance of 2.31 between the vegetation point cloud and the cable point cloud. This is below the safe distance threshold of 4.0 for this voltage level. For the points exceeding the limit, perform density-based spatial clustering with a neighborhood radius of 2. The minimum sample size is 10, and the clustering results are merged into one potential hazard area, such as... Figure 3 As shown.

[0079] In step S06, the multi-factor weighted comprehensive risk assessment model integrates the cable spacing parameters and structural deformation residuals of the span. and the area of ​​vegetation hazard zone, 220 Voltage level correction factor, 32℃ ambient temperature correction factor and 8 The wind speed correction factor was used to dynamically adjust the safety distance threshold, ultimately classifying the vegetation hazard at this location as high-risk. The risk assessment results for all 168 sections of the entire corridor are summarized in Table 3.

[0080] Table 3. Statistical Table of Risk Level Distribution Across All Channels

[0081] In an interactive 3D visualization scene built on WebGL technology, high-risk and extremely high-risk levels are marked with dynamic warning symbols, and structured warning reports are pushed to the inspection management system in real time through the application programming interface.

[0082] Compared to traditional manual inspections and methods relying solely on geometric thresholds, this invention represents a significant technological advancement. Traditional manual inspections depend on visual judgment by inspectors, failing to quantitatively extract tower deformation parameters and cable sag deformation components, and also unable to distinguish between normal physical deformation and abnormal structural deformation. Simple geometric thresholding only calculates distances to the point cloud, lacking the ability to complete missing point clouds. In severely obstructed spans, this can easily lead to missed hazards or misjudgments of normal structures as deformations due to missing point clouds. This invention, by extracting global manifold invariants in the frequency domain, eliminates reliance on local neighborhood features during the completion process, addressing the fundamental flaw of spatial domain methods—feature fragmentation due to local deficiencies—from a frequency domain analysis perspective of signal processing. Furthermore, by combining the physical force equations of the catenary with the spatial variational principle, it mathematically decouples normal sag changes caused by temperature from abnormal structural deformation, preventing the misjudgment of normal sag changes as line faults. This provides hazard identification with a physical mechanism support rather than purely geometric comparison.

[0083] like Figure 3 As shown, the vegetation hazard clustering areas are marked with high-risk dynamic symbols in the 3D visualization scene, clearly demonstrating the spatial distance relationship between the vegetation point cloud and the cable point cloud, as well as the hazard clustering range.

[0084] It should be noted that the variables involved in this invention are explained in detail in Tables 4 and 5.

[0085] Table 4. Variable Explanation Table (Part 1)

[0086] Table 5. Variable Explanation Table (Part Two)

[0087] The above description is merely a specific 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 intelligent identification and early warning of hidden dangers in power transmission lines, characterized in that, Includes the following steps: Three-dimensional point cloud data of power transmission channels are collected by a mobile platform equipped with laser scanning equipment. A distributed external storage dual-layer decoupled index is constructed. The channel is divided into geographical data units according to the span and tower location. The upper layer adopts a lightweight spatial grid index for macro topology, and the lower layer dynamically swaps in and out local point cloud files based on memory mapping technology. A multi-pulse echo full waveform spatiotemporal correlation separation algorithm is applied to the point cloud within the geographic data unit. The laser echo amplitude, pulse half-width, full width at half-height, and time series energy distribution are jointly analyzed to construct a single-point physical property scattering matrix, remove secondary scattering artifact points, and extract tower point clouds and cable point clouds. For the incomplete tower point cloud and fitting point cloud, the curvature flow modulation fractal progressive reconstruction model is input to complete the missing regions and output the complete tower point cloud. During the completion process, the spectral tensor kernel width parameter is dynamically given by the density adaptive adjustment function based on the point density estimate of the local region, the normal vector consistency index and the feature difference degree of adjacent levels. An enhanced point cloud matching framework based on feature descriptors is adopted to perform three-layer pyramid multi-resolution iterative nearest point registration with the benchmark design point cloud to obtain an accurate spatial transformation matrix. Based on the spatial transformation matrix, the tower translation vector and rotation angle are extracted. The catenary physical force equation and the spatial variation principle cable sag deformation decoupling algorithm are applied to the cable point cloud. The deformation residual is quantitatively output as the hidden danger feature. Density-based spatial clustering is performed on the over-limit points to merge the hidden danger areas. A multi-factor weighted comprehensive risk assessment model was established, extracting tower deformation parameters, cable spacing parameters, and characteristic parameters of potential hazard areas. Voltage level, ambient temperature, and wind speed correction coefficients were introduced to dynamically adjust the safe distance threshold, and the risk level was divided into four levels. An interactive three-dimensional visualization scene was constructed based on WebGL technology, generating a structured early warning report and pushing it to the inspection management system.

2. The intelligent identification and early warning method for hidden dangers in transmission lines according to claim 1, characterized in that, The distributed external storage dual-layer decoupled index specifically divides the power transmission channel into independent geographic data units with high cohesion and low coupling in geographic space. The upper layer is a lightweight spatial grid that stores the spatial boundary bounding box coordinates and file pointers of each geographic data unit. The lower layer uses memory mapping technology to directly map the disk address space of the point cloud file to the virtual address space of the process, allowing swapping in and out as needed. The computing power of each geographic data unit is decoupled, and multi-processor parallel scheduling is supported.

3. The intelligent identification and early warning method for hidden dangers in transmission lines according to claim 2, characterized in that, The multi-pulse echo full waveform spatiotemporal correlation separation algorithm specifically performs Gaussian decomposition on the full waveform digital record, fits the synthesized echo into the sum of several single-peak Gaussian components, establishes a single-point physical property scattering matrix for each scattering event, introduces spatiotemporal coherence constraints to eliminate secondary scattering artifacts, and assigns the point cloud to the corresponding physical target based on the time and amplitude of each scattering component.

4. The intelligent identification and early warning method for hidden dangers in transmission lines according to claim 3, characterized in that, In the Gaussian decomposition, the upper limit of the number of components is 6. This value is determined by performing a full waveform scan of the parallel steel cable in a laboratory environment and using the Akaike Information Criterion as the basis for model selection.

5. The intelligent identification and early warning method for hidden dangers in transmission lines according to claim 4, characterized in that, The curvature flow modulation fractal progressive reconstruction model is specifically composed of a spectral encoder and a three-stage fractal progressive growth decoder connected by three layers of cross-stage jumps. The spectral encoder consists of three cascaded spectral neural network layers, which extract the global manifold invariants of the incomplete topology in the frequency domain. The three-stage fractal progressive growth decoder dynamically adjusts the splitting factor with curvature as a spatial feedback signal. The adversarial multi-scale discriminator simultaneously evaluates the generated point cloud at both the global topological consistency scale and the local microscopic geometric texture scale.

6. The intelligent identification and early warning method for hidden dangers in transmission lines according to claim 5, characterized in that, The spectral graph neural network layer specifically uses the K-nearest neighbor algorithm to construct a point cloud graph structure, performs Chebyshev polynomial approximate spectral convolution on the graph Laplacian matrix in the frequency domain, and implements step-by-step downsampling by sampling the farthest point after each layer. The graph convolution kernel weight matrix adopts a row-level sparse allocation strategy to allocate thread bundles proportionally according to the node degree.

7. The intelligent identification and early warning method for hidden dangers in transmission lines according to claim 6, characterized in that, The enhanced point cloud matching framework employs a three-layer pyramid multi-resolution iterative nearest-point registration. Specifically, layer one performs fast coarse registration on point clouds with a retention ratio equal to the coarse registration retention value; layer two performs medium-precision registration on point clouds with a medium retention ratio; and layer three performs high-precision fine registration on the complete point cloud. The registration result of each layer serves as the initial value for the next layer. The hybrid distance metric adaptively adjusts the weights of point-to-point distance and point-to-surface distance based on local curvature characteristics.

8. The intelligent identification and early warning method for hidden dangers in transmission lines according to claim 7, characterized in that, In the three-layer pyramid multi-resolution iterative nearest point registration, the retention ratio of point cloud in the first layer is 1%, the retention ratio of point cloud in the second layer is 10%, and the retention ratio of point cloud in the third layer is a complete point cloud. The retention ratio of each layer is determined iteratively by using the registration success rate and convergence speed as evaluation indicators in the registration experiments with different initial pose deviations.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores program instructions, which, when executed in a computer, are used to perform the intelligent identification and early warning method for hidden dangers in power transmission lines as described in any one of claims 1-8.

10. A smart identification and early warning system for hidden dangers in power transmission lines, characterized in that, The system comprises the computer-readable storage medium of claim 9, wherein the system is a computer, the computer-readable storage medium is disposed within the system, and the system is provided with a microprocessor that executes program instructions stored in the computer-readable storage medium.

Images