A power transmission channel hidden danger real-time three-dimensional reconstruction and ranging method based on a fixed laser radar, medium and system

Abstract

The application provides a power transmission channel hidden danger real-time three-dimensional reconstruction and ranging method, medium and system based on a fixed laser radar, belongs to the technical field of power transmission channels, and collects three-dimensional point clouds of a power transmission corridor through the fixed laser radar, carries out adaptive segmentation and multi-stage filtering on the point clouds according to local geometric feature entropy and density unevenness, predicts a conductor deformation trend through an artificial intelligence bidirectional feedback space-time deformation residual model, calculates the minimum space interval between an intruder and the conductor by using an adaptive space dynamic bounding box ranging algorithm driven by an electrical geometric model, and drives graded warning according to a safety situation level, so that the technical problems that the fixed laser radar cannot realize real-time collaborative reconstruction of conductor geometric state and environment coupling deformation in a complex dynamic scene of the power transmission channel and the ranging precision significantly declines with changes in the environment and load are solved.

Description

Technical Field

[0001] This invention belongs to the field of power transmission channel technology. Specifically, it relates to a method, medium, and system for real-time three-dimensional reconstruction and ranging of potential hazards in power transmission channels based on fixed lidar. Background Technology

[0002] In the field of power transmission corridor hazard detection, traditional methods rely on manual inspections, aerial photography, or mobile lidar scanning to identify hazards by collecting static geometric information of the corridor environment and conducting manual visual assessment. With the widespread adoption of lidar technology, fixed lidar has been introduced into power transmission corridor monitoring scenarios. It can continuously collect three-dimensional point clouds of intruders such as trees and buildings within the corridor, and combine this with static ranging algorithms such as Euclidean distance to calculate the spatial distance between the intruder and the conductor. However, in current fixed lidar power transmission corridor monitoring, the actual sag and spatial attitude of the conductor continuously change dynamically under different current loads, ambient temperatures, solar radiation intensity, and wind speeds. Static point cloud reconstruction methods fail to couple the thermodynamic physical state with the point cloud geometric reconstruction process, resulting in significant deviations between the actual sag of the conductor under high load, high temperature, or strong wind conditions and the ranging reference. Furthermore, blind spots cause incomplete point cloud geometry, further leading to discontinuities in the reconstructed surface. In other words, existing technologies have the technical problem that fixed lidar cannot achieve real-time collaborative reconstruction of the geometric state of the conductor and the coupled deformation of the environment in complex dynamic scenarios of power transmission channels, resulting in a significant decline in ranging accuracy as the environment and load change. Summary of the Invention

[0003] In view of this, the present invention provides a method, medium and system for real-time three-dimensional reconstruction and ranging of hidden dangers in power transmission channels based on fixed lidar, which can solve the technical problem in the prior art that fixed lidar cannot achieve real-time collaborative reconstruction of the geometric state of the conductor and the coupled deformation of the environment in complex dynamic scenarios of power transmission channels, resulting in a significant decline in ranging accuracy with changes in environment and load.

[0004] The present invention is implemented as follows: The first aspect of the present invention provides a method for real-time three-dimensional reconstruction and ranging of potential hazards in power transmission channels based on fixed lidar, comprising the following steps: A fixed lidar sensing unit is installed on the top of the transmission tower to perform a panoramic scan of the transmission corridor area and collect an initial three-dimensional point cloud dataset. By integrating local geometric feature entropy and density non-uniformity, the initial 3D point cloud dataset is adaptively segmented, separating vegetation points and power grid equipment and facility points, and then filtering out redundant points, discrete noise points and high-frequency disturbance noise in sequence to output clean point cloud data. For the candidate point set of the conductor in the clean point cloud data, the catenary physical constraint cutting plane reconstruction algorithm is used to remove impurity points that deviate from the physical trajectory, and the three-dimensional trajectory of the conductor is obtained. By capturing local morphological features of clean point cloud data through hierarchical feature learning, a neural implicit network is driven to regress the directed distance field. Blind zone matching and completion are performed by combining curvature continuity constraints. The moving cube algorithm and non-uniform rational B-splines are used to mesh and fit the implicit geometric expression to the surface, and a continuous channel three-dimensional surface model is output. Intruders or abnormal objects are identified in the three-dimensional surface model of the channel. A physical-thermodynamic coupling model of the conductor state is established. The theoretical sag correction is calculated based on the current load, ambient temperature, solar radiation intensity and wind speed. An adaptive spatial dynamic bounding box ranging algorithm driven by the electrical geometry model is used. Combined with the deformation trend output by the bidirectional feedback spatiotemporal deformation residual model, the minimum spatial interval between the intruder and the conductor is measured. Based on the minimum spatial interval and security status level, the alarm system is driven to issue graded warnings and generate handling suggestions.

[0005] The local geometric feature entropy refers to the statistical dispersion of the reflection intensity, normal vector angle, and curvature distribution within the local neighborhood of the point cloud; the density non-uniformity refers to the statistical quantity of the difference in density of the point cloud distribution per unit space.

[0006] Specifically, the adaptive segmentation involves inputting the local geometric feature entropy and density non-uniformity into a classification framework composed of region growing and graph cut theory. Region growing gradually expands the set of similar points based on neighborhood similarity, while graph cut transforms the segmentation problem into a minimum cut problem in graph theory. The combination of the two decouples the power grid equipment and facility points from the vegetation points.

[0007] Specifically, the sequential filtering of redundant points, discrete noise points, and high-frequency disturbance noise involves the following steps: Redundant points are filtered using spatial grid partitioning and point distance analysis, fusing points that fall within the same micro-element grid or whose Euclidean distance to neighboring points is less than a distance threshold; discrete noise points are filtered using an adaptive parameter-based density clustering algorithm, which accelerates neighborhood queries through a spatial index structure and dynamically determines the clustering radius and minimum point threshold based on the point cloud distribution characteristics; and high-frequency disturbance noise is filtered using a feature-preserving smoothing algorithm, which distinguishes between flat regions and feature edges based on local curvature and normal vectors, and uses a differentiated smoothing intensity factor for filtering.

[0008] Specifically, the catenary physical constraint cutting plane reconstruction algorithm is used to construct a three-dimensional nonlinear catenary equation. ,in , , The parameters are the morphological parameters of the catenary. In each iteration of the algorithm, a convex hull cutting plane is adaptively constructed based on the Euclidean distance from the point cloud to the currently fitted catenary trajectory. Impurity points whose distance exceeds the threshold of the cutting plane are gradually cut off and separated. The optimal catenary parameters are solved iteratively by the Newton-Raphson method.

[0009] Specifically, the neural implicit network regresses the directed distance field by encoding local morphological features into local context prior parameters. The neural query network predicts the offset of the spatial query point relative to the potential surface based on the local context prior parameters. The offset query point position and the local context prior parameters are then input into the neural implicit function to jointly predict the directed distance value of the point.

[0010] The neural query network and the neural implicit function are trained end-to-end through a joint loss function. The joint loss function is composed of a weighted sum of spatial projection loss, directed distance field regression loss, and implicit surface constraint loss. The spatial projection loss calculates the square of the Euclidean distance between the offset query point and the nearest point on the real surface. The directed distance field regression loss calculates the square of the difference between the predicted directed distance value and the real directed distance value. The implicit surface constraint loss requires the gradient magnitude of the directed distance field at the query point to be close to 1.

[0011] Specifically, the moving cube algorithm involves spatially dividing the global directed distance field using a regular three-dimensional mesh, querying the directed distance field sign and value at the eight vertices of the voxel, determining the topology of triangular facet connections through a lookup table based on the isosurface threshold, and calculating the vertex coordinates of the facets through linear interpolation; for traverse regions with gentle gradient changes, a multi-resolution subdivision strategy is used to generate a high-fidelity triangular mesh.

[0012] The physical-thermodynamic coupling model of the conductor state is established by incorporating the Joule heat generated by the current load, the ambient temperature, the heat absorbed by solar radiation and the heat dissipation by wind speed into a unified thermal balance equation for coupling and solving, thereby obtaining the conductor temperature change, and then calculating the sag correction based on the thermal expansion and contraction coefficient of the conductor material.

[0013] The bidirectional feedback spatiotemporal deformation residual model is composed of a three-dimensional spatial geometry extraction branch and a time series long short-term memory branch in parallel. Cross-layer jump connections and reverse correction loops are established between residual blocks. The deformation acceleration trend output by the time series branch is used as a gating signal to reversely fine-tune the weights of the preceding spatial features. The catenary physical equation is introduced into the loss function as a regularization constraint layer.

[0014] Specifically, the adaptive spatial dynamic bounding box ranging algorithm driven by the electrical geometry model calculates the irregular spatial safety conduit bounding box around the axis of the non-uniform rational B-spline curve of the conductor based on the dynamic electrical insulation distance and wind deflection angle corresponding to the conductor voltage level. The intruder point cloud is spatially partitioned using a hierarchical bounding box tree. Coarse screening is performed by judging whether the intruder bounding box intersects with the conductor electrical conduit bounding box. After intersection, the local shortest path graph theory algorithm is activated to finely solve for the minimum dynamic three-dimensional spatial distance.

[0015] The security status level refers to the discretized level obtained by performing spatiotemporal superposition analysis on the spatial attitude of the conductor output by the physical-thermodynamic coupling model of the intruder's predicted trajectory and the conductor's state, and then combining the spatial intersection probability with the current minimum spatial interval value.

[0016] The bidirectional feedback spatiotemporal deformation residual model also includes a resource adjustment function, which calculates the resource adjustment value based on the current batch point cloud data volume, the number of iteration layers of the network residual block, and the number of activated cross-layer skip connections. Based on resource adjustment value The interval dynamic adjustment strategy for video memory allocation and computing flow allocation.

[0017] Among them, resource adjustment value The following allocation rules apply: when When in the low range, a single computation stream is used to serially execute the calculation of each residual block; when When in the middle range, a dual computational flow is used to compute the spatial geometry extraction branch and the time series short-term memory branch in parallel; when When in the high range, multiple computational flows are used to dynamically divide the number of flows based on the number of cross-layer jump connections, and the memory regions corresponding to the reverse correction loop are allocated independently.

[0018] Among them, resource adjustment value The lower range is The middle section is The high range is The weight coefficients of spatial projection loss, directed distance field regression loss, and implicit surface constraint loss in the joint loss function were obtained through multiple training experiments on the validation set and combined with convergence effect analysis; catenary morphology parameters , , The initial value range was obtained by conducting multiple curve fitting experiments on historical traverse point cloud data and combining them with least squares error analysis, and then determined by iterative correction using field measured data.

[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 real-time three-dimensional reconstruction and ranging of potential hazards in power transmission channels based on fixed lidar.

[0020] A third aspect of the present invention provides a real-time three-dimensional reconstruction and ranging system for potential hazards in power transmission channels based on fixed lidar, 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 establishes a physical-thermodynamic coupling model of the conductor state, incorporating current load, ambient temperature, solar radiation intensity, and wind speed into a unified thermal balance equation. This allows for real-time correction of conductor sag and spatial attitude. Simultaneously, a neural implicit network regresses the directed distance field and combines curvature continuity constraints to fill in obstructed blind spots, constructing a continuous three-dimensional surface model of the channel. Furthermore, a bidirectional feedback spatiotemporal deformation residual model is used to fuse temporal point cloud data with multi-source sensor data to predict deformation trends. This provides a dynamic geometric reference that consistently corresponds to the actual physical state of the conductor for the adaptive spatial dynamic bounding box ranging algorithm driven by the electrical geometry model. This ensures that the ranging results remain consistent with the actual spatial position of the conductor even as environmental and load conditions such as current, temperature, and wind speed continuously change. This solves the technical problem of ranging accuracy degradation caused by the static reconstruction reference deviating from the actual conductor sag. In summary, this invention addresses the technical problem mentioned in the background art: fixed lidar cannot achieve real-time collaborative reconstruction of the conductor's geometric state and environmental coupling deformation in complex dynamic scenarios of power transmission channels, leading to a significant degradation in ranging accuracy with changes in environment and load. Attached Figure Description

[0022] Figure 1 This is a flowchart of the method of the present invention.

[0023] Figure 2 A comparison chart of conductor sag corrections under different operating conditions.

[0024] Figure 3 This is a curve showing the iterative convergence process of the bidirectional feedback spatiotemporal deformation residual model. Detailed Implementation

[0025] 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.

[0026] like Figure 1 The diagram shown is a flowchart of a method for real-time three-dimensional reconstruction and ranging of hidden dangers in power transmission channels based on fixed lidar, provided by the first aspect of this invention. This method includes the following steps: S1. Install a fixed lidar sensing unit on the top of the transmission tower to perform a panoramic scan of the transmission corridor area and collect an initial three-dimensional point cloud dataset. S2. By fusing local geometric feature entropy and density non-uniformity, the initial three-dimensional point cloud dataset is adaptively segmented to separate vegetation points and power grid equipment and facility points, and redundant points, discrete noise points and high-frequency disturbance noise are filtered out in sequence to output clean point cloud data. S3. For the candidate point set of the conductor in the clean point cloud data, the catenary physical constraint cutting plane reconstruction algorithm is used to remove impurity points that deviate from the physical trajectory, and the three-dimensional trajectory of the conductor is obtained. S4. By learning hierarchical features, the local morphological features of the clean point cloud data are captured, the neural implicit network is driven to regress the directed distance field, and blind zone matching and completion are performed by combining curvature continuity constraints. The moving cube algorithm and non-uniform rational B-splines are used to mesh and fit the implicit geometric expression to the surface, and a continuous channel three-dimensional surface model is output. S5. Identify intruders or abnormal objects in the three-dimensional surface model of the channel, establish a physical-thermodynamic coupling model of the conductor state, calculate the theoretical sag correction based on current load, ambient temperature, solar radiation intensity and wind speed, and use an adaptive spatial dynamic bounding box ranging algorithm driven by the electrical geometry model, combined with the deformation trend output by the bidirectional feedback spatiotemporal deformation residual model, to measure the minimum spatial interval between the intruder and the conductor. S6. Based on the minimum spatial interval and security status level, drive the alarm system to execute graded warnings and generate handling suggestions; The local geometric feature entropy refers to the statistical dispersion of the reflection intensity, normal vector angle, and curvature distribution within the local neighborhood of the point cloud. It is used to characterize the irregularity of the local structure of the point cloud. The higher the value, the more irregular the geometric shape of the region, which is mostly caused by scattered reflections from vegetation. The density non-uniformity refers to the statistical quantity of the difference in the density of the point cloud distribution per unit space. The two are combined and input into the classification framework composed of region growing and graph cut theory. Region growing gradually expands the set of similar points based on neighborhood similarity. Graph cut transforms the segmentation problem into the minimum cut problem in graph theory. The combination of the two is used to finely decouple the power grid equipment and facility points from the vegetation points. The specific methods for filtering redundant points are as follows: based on spatial grid division and point distance analysis, points falling within the same micro-element grid or whose Euclidean distance to neighboring points is less than a preset threshold are merged; the specific methods for filtering discrete noise points are as follows: using an adaptive parameter-based density clustering algorithm, accelerating neighborhood queries through a spatial index structure, and dynamically determining the clustering radius and minimum point number threshold based on the point cloud distribution characteristics; the specific methods for filtering high-frequency disturbance noise are as follows: applying a feature-preserving smoothing algorithm, distinguishing flat regions from feature edges based on local curvature and normal vectors, and using a differentiated smoothing intensity factor for filtering. The distance threshold, clustering radius, and smoothing intensity factor values ​​involved in the above filtering process were obtained through multiple filtering experiments on multiple sets of typical scene point cloud samples, combined with statistical analysis of surface errors before and after filtering.

[0027] The catenary physical constraint cutting plane reconstruction algorithm is based on the physical law that the conductor obeys the catenary equation under its own weight, and constructs a three-dimensional nonlinear catenary equation: In the formula , , The parameters for the catenary morphology represent curve tension, horizontal offset, and vertical offset, respectively. Their initial values ​​were obtained through multiple curve fitting experiments on historical traverse point cloud data combined with least squares error analysis, and then iteratively corrected using field measurement data. In each iteration, the algorithm adaptively constructs a convex hull cutting plane based on the Euclidean distance from the point cloud to the currently fitted catenary trajectory. Tree and bird points exceeding a certain distance are gradually clipped and separated. This threshold is obtained through experimental statistical analysis of multiple typical scene point clouds, and the optimal catenary parameters are iteratively solved using the Newton-Raphson method. This algorithm introduces physical equation constraints into the point cloud reconstruction process, giving the noise clipping process a physically sound basis, avoiding misjudgments caused by relying solely on statistical thresholds, improving the continuity of traverse trajectory reconstruction and the reliability of the ranging benchmark, and providing a stable geometric benchmark for subsequent dynamic ranging.

[0028] The specific method of the neural implicit network regression of the directed distance field is as follows: the local morphological features are encoded as local context prior parameters, and the offset of the spatial query point relative to the potential surface is predicted by the neural query network based on the local context prior parameters. The offset query point position and the local context prior parameters are input together into the neural implicit function to jointly predict the directed distance value of the point. The neural query network and the neural implicit function are trained end-to-end through a joint loss function. The joint loss function is composed of a weighted sum of spatial projection loss, directed distance field regression loss, and implicit surface constraint loss. The spatial projection loss calculates the square of the Euclidean distance between the offset query point and the nearest point on the real surface. The directed distance field regression loss calculates the square of the difference between the predicted directed distance value and the real directed distance value. The implicit surface constraint loss requires that the gradient magnitude of the directed distance field at the query point is close to 1. The weight coefficients of the three losses are obtained by conducting multiple training experiments on the validation set and combining convergence effect analysis. The moving cube algorithm spatially divides the global directed distance field using a regular 3D mesh. It queries the directed distance field sign and value at the eight vertices of a voxel, determines the topology of triangular facet connections using a lookup table based on a preset isosurface threshold, and calculates the vertex coordinates of the facets using linear interpolation. For conductor regions with gentle gradient changes, a multi-resolution subdivision strategy is used to generate a high-fidelity triangular mesh. The non-uniform rational B-spline, a commonly used curve and surface representation method in computer graphics, expresses the surface morphology through control points and weight factors. It is used to adaptively fit and complete the broken conductors and tower blind spots in the high-fidelity triangular mesh, ensuring geometric continuity of the reconstructed scene and avoiding ranging errors caused by holes. The historically complete 3D point cloud prior model refers to a benchmark model established from 3D point cloud data collected from multiple historical periods of the same transmission channel and manually verified as complete and unobstructed. It is used to provide morphological matching references for blind spot areas.

[0029] The aforementioned physical-thermodynamic coupling model of the conductor state is established based on the principle of conductor thermal balance. It incorporates Joule heat generated by the current load, ambient temperature, solar radiation absorption heat, and wind speed heat dissipation into a unified thermal balance equation for coupling and solution, yielding the conductor temperature change. The sag correction is then calculated based on the conductor material's thermal expansion and contraction coefficient. The coefficients in the equation are obtained through laboratory thermal experiments on the conductor material and regression analysis using long-term field monitoring data. This model transforms the conductor's dynamic thermodynamic state into a quantifiable sag correction factor, enabling the conductor's geometric position in the reconstructed space to adjust in real time according to environmental and load changes. This avoids false alarms or missed alarms caused by static ranging at different times, improving the stability of all-weather ranging.

[0030] The adaptive spatial dynamic bounding box ranging algorithm driven by the electrical geometry model combines electrical insulation safety distance standards with computer graphics bounding box technology. Based on the dynamic electrical insulation distance corresponding to the conductor voltage level and the wind deflection angle, it calculates and forms an irregularly shaped spatial safety conduit bounding box around the conductor's B-spline curve axis. The dynamic electrical insulation distance value is obtained based on relevant power industry safety specifications and corrected using on-site meteorological experimental data. The algorithm uses a hierarchical bounding box tree to spatially partition the intrusion point cloud. By determining whether the intrusion bounding box intersects with the conductor's electrical conduit bounding box, it avoids the computational resource consumption caused by calculating the distance point by point across the entire scene. Once spatial intersection occurs, a local Dijkstra's shortest path graph theory algorithm is activated to solve for the shortest dynamic three-dimensional spatial distance from the intrusion surface to the conductor axis within the intersecting subspace. This algorithm, through a combination of hierarchical partitioning and local fine-grained solution, significantly reduces the resource consumption of the global point cloud while ensuring ranging accuracy, improving real-time response capabilities in scenarios with large data volumes.

[0031] The bidirectional feedback spatiotemporal deformation residual model is a deep learning-based model for predicting the dynamic deformation of conductors. The associated data includes serialized time-series 3D point cloud spatial coordinates, lidar ranging timestamps, and time-series data from environmental wind speed and temperature sensors. The model's specific structure is as follows: the network consists of a parallel 3D spatial geometry extraction branch and a time-series long short-term memory branch. The spatial geometry extraction branch uses a multi-layer convolutional structure to abstract features from the serialized time-series 3D point cloud spatial coordinates layer by layer, extracting the geometric morphological features of the conductor in 3D space. The time-series long short-term memory branch receives the feature sequence output from the spatial branch and, combined with lidar ranging timestamps and environmental wind speed and temperature sensor time-series data, models the trend of conductor deformation evolution over time. The core improvement of the network lies in establishing cross-layer jump connections and reverse correction loops between residual blocks, forming a bidirectional internal iterative feedback mechanism. After the spatial feature input activates the time-series branch, the deformation acceleration trend output by the time-series branch serves as a gating signal to fine-tune the weights of the previous few layers of spatial features, forming an iterative process of alternating forward propagation and reverse correction. The network internally sets three iterative cycles to gradually approximate the deformation reconstruction result. Meanwhile, the network incorporates the physical equation of the catenary in high-voltage transmission lines as a regularization constraint layer in the loss function. This incorporates the deviation between the network's output shape and the theoretical shape of the catenary equation into the loss calculation, ensuring that the feature extraction process is both data-driven and constrained by physical laws. The training dataset for the model is established as follows: historical point cloud sequence data of multiple transmission channels under different seasons, loads, and meteorological conditions are collected, along with corresponding data on current load, ambient temperature, solar radiation intensity, and wind speed. The point cloud data is manually labeled with the actual deformation variables as tags, and then divided into training sample pairs according to the time series. The model training steps are as follows: the training sample pairs are input into the network, and the sum of the predicted deformation variables and the catenary physical constraint loss is calculated forward as the total loss. The network weights are updated through backpropagation, and this iteration is repeated until the loss converges. An independent validation set is used to evaluate and adjust the model's generalization ability. This model provides a basis for predicting deformation trends throughout the reconstruction process, allowing spatial geometric features and temporal evolution trends to mutually correct each other during iteration. This results in a 3D deformation reconstruction result that more closely approximates the actual physical bending shape of the conductor, improving the continuity and accuracy of 3D reconstruction in dynamic scenarios.

[0032] The resource adjustment function is used to adjust the GPU memory allocation and CUDA stream allocation of the bidirectional feedback spatiotemporal deformation residual model during training and inference. The function performs a comprehensive calculation based on the current batch point cloud data volume, the number of iteration layers of the network residual block, and the number of cross-layer skip connection activations to obtain a resource adjustment value, denoted as . .when At that time, a single CUDA stream is used to serially execute the calculation of each residual block, and the memory allocation is limited to a lower range; when At that time, a dual CUDA stream parallel allocation is adopted, where the spatial geometry extraction branch and the time series long short-term memory branch are allocated to different streams for parallel computation, and the memory allocation is divided according to the proportion of the branch feature dimensions; when At the same time, a multi-CUDA stream allocation method is adopted, dynamically dividing the number of streams based on the number of cross-layer skip connections, and independently allocating the memory area corresponding to the reverse correction loop to avoid memory contention between the reverse correction process and the forward computation process. The specific calculation method of the resource adjustment value and the threshold of each interval were obtained by conducting multiple training and inference experiments on point cloud data of different scales in an experimental environment, and statistically analyzing the relationship between resource usage and computation time.

[0033] The security status level refers to the discretized level obtained by performing spatiotemporal superposition analysis on the spatial attitude of the conductor output by the physical-thermodynamic coupling model of the intruder's predicted trajectory and the conductor's state, combined with the spatial intersection probability and the current minimum spatial interval value. The spatial intersection probability is obtained by performing multiple probability simulations and statistics on the predicted trajectory and the conductor's attitude. The level classification threshold is determined by statistical analysis of historical hidden danger cases and multiple experimental verifications.

[0034] The specific implementation of step S1 is as follows: A fixed lidar sensing unit is installed on the top of the transmission tower. Utilizing the tower's high vantage point, a comprehensive scan of the transmission corridor area is performed, collecting an initial 3D point cloud dataset including vegetation, buildings, towers, and conductors within the corridor. During the scan, the lidar emits laser pulses sequentially in all azimuth and elevation directions at a fixed angular resolution and receives the echoes. Each echo records its corresponding 3D spatial coordinates and reflection intensity value, forming a dense set of spatial points. The initial 3D point cloud dataset contains mixed echo information of various terrain features, providing the initial data foundation for subsequent segmentation and cleanup.

[0035] The specific implementation of step S2 is as follows: First, calculate the statistical dispersion of reflection intensity, normal vector angle, and curvature distribution within the local neighborhood of each point to obtain the local geometric feature entropy; simultaneously, statistically analyze the density differences of the point cloud distribution within a unit space to obtain the density non-uniformity. The above two features are then jointly input into a classification framework composed of region growing and graph cut theory. Region growing starts from the seed point and gradually expands the set of similar points based on neighborhood geometric similarity. Graph cut transforms the segmentation problem into a minimum cut problem in graph theory. The two work together to finely decouple vegetation points from power grid equipment and facility points, completing adaptive segmentation. After segmentation, a three-level filtering process is performed sequentially: The first level filters out redundant points. Based on spatial grid partitioning, the point cloud space is divided into several micro-grids. Points falling within the same micro-grid or whose Euclidean distance to neighboring points is less than a distance threshold are fused. The distance threshold reference value is 5–20. The second stage filters out discrete noise points using an adaptive parameter-based density clustering algorithm. A spatial index structure accelerates neighborhood lookups, and the clustering radius and minimum point threshold are dynamically determined based on the point cloud distribution characteristics. The reference value for the clustering radius is 0.1–0.5. The third stage filters out high-frequency disturbance noise and applies a feature-preserving smoothing algorithm. Based on local curvature and normal vectors, it automatically distinguishes between flat regions and feature edges, applies a stronger smoothing intensity factor to flat regions and a weaker smoothing intensity factor to feature edges, and outputs clean point cloud data.

[0036] The specific implementation of step S3 is as follows: Extract candidate points for the conductor from the clean point cloud data based on local curvature and neighborhood connectivity, and construct the three-dimensional spatial nonlinear catenary equation. ,in , , These represent curve tension, horizontal offset, and vertical offset, respectively. The initial values ​​were determined through multiple curve fitting experiments using historical traverse point cloud data and iteratively corrected using field measurement data. In each iteration, the algorithm adaptively constructs a convex hull cutting plane based on the Euclidean distance from each point in the candidate point set to the currently fitted catenary trajectory. Tree and bird points whose distances exceed the cutting plane threshold are progressively trimmed and separated. The cutting plane threshold reference value is 0.3–0.8. The optimal catenary parameters are updated iteratively using the Newton-Raphson method. , , This continues until the residual converges, yielding a three-dimensional trajectory of the conductor that conforms to physical laws.

[0037] The specific implementation of step S4 is as follows: A multi-layer convolutional structure is used to perform hierarchical feature learning on the clean point cloud data, extracting local morphological features layer by layer, and encoding the extracted features as local context prior parameters. The neural query network predicts the offset of the spatial query point relative to the potential surface based on the local context prior parameters. The offset query point position and the local context prior parameters are input together into the neural implicit function to predict the directed distance value of the point, forming a global directed distance field. The neural query network and the neural implicit function are trained end-to-end using a joint loss function. The joint loss function consists of a weighted sum of spatial projection loss, directed distance field regression loss, and implicit surface constraint loss. The weight coefficients of the three losses are obtained through multiple training experiments on the validation set and combined with convergence effect analysis, with a reference weight ratio of 1:1:0.1. For blind areas, a morphological matching reference is provided by combining a historically complete 3D point cloud prior model, and curvature continuity constraints are applied for matching and completion. The moving cube algorithm spatially divides the global directed distance field using a regular 3D mesh. It queries the directed distance field sign and value at the eight vertices of the voxel, determines the topology of the triangular facets by looking up a lookup table based on the isosurface threshold, and calculates the vertex coordinates of the facets by linear interpolation. A multi-resolution subdivision strategy is used to generate a high-fidelity triangular mesh for the conductor region. Finally, non-uniform rational B-splines are used to adaptively fit and complete the broken conductor and tower blind area, outputting a geometrically continuous 3D surface model of the channel.

[0038] The specific implementation of step S5 is as follows: Intruders or abnormal objects are identified in the three-dimensional surface model of the channel. A physical-thermodynamic coupling model of the conductor state is established based on the conductor thermal balance principle. Joule heat generated by the current load, ambient temperature, solar radiation absorption heat, and wind speed heat dissipation are incorporated into a unified thermal balance equation for coupling and solution to obtain the conductor temperature change. This change is then converted into a sag correction amount using the conductor material's thermal expansion and contraction coefficient, and the geometric position of the conductor in the reconstructed space is adjusted in real time. The bidirectional feedback spatiotemporal deformation residual model receives serialized time-series three-dimensional point cloud spatial coordinates, lidar ranging timestamps, and environmental wind speed and temperature sensor time-series data. The three-dimensional spatial geometry extraction branch extracts the conductor's spatial geometric morphology features, and the time-series long-short-term memory branch models the deformation evolution trend over time. Cross-layer jump connections and reverse correction loops between residual blocks constitute a bidirectional iterative feedback mechanism. The deformation acceleration trend output by the time-series branch is used as a gating signal to fine-tune the weights of the preceding spatial features. The network internally sets the number of iterations to 3. The catenary physical equation is introduced into the loss function as a regularization constraint layer, outputting the deformation trend. An adaptive spatial dynamic bounding box ranging algorithm driven by an electrical geometry model constructs an irregular spatial safety conduit bounding box around the axis of a non-uniform rational B-spline curve of the conductor, based on the dynamic electrical insulation distance and wind deflection angle corresponding to the conductor voltage level. The intruder point cloud is spatially partitioned using a hierarchical bounding box tree. The algorithm determines whether the intruder bounding box intersects with the conductor electrical conduit bounding box. If they intersect, the algorithm activates a local shortest path graph theory algorithm to finely solve for the minimum dynamic three-dimensional spatial distance in the intersecting subspace, thus measuring the minimum spatial interval between the intruder and the conductor.

[0039] The specific implementation of step S6 is as follows: Based on the spatial attitude of the conductor output by the physical-thermodynamic coupling model of the intruder's predicted trajectory and the conductor's state, a spatiotemporal superposition analysis is performed. The safety status level is then determined by combining the spatial intersection probability and the minimum spatial interval value. The level classification threshold is determined through statistical analysis of historical hazard cases and multiple experimental verifications. The alarm system is driven to execute graded warnings based on the safety status level. Simultaneously, handling suggestions are generated based on the intruder type and the current channel geometry for maintenance personnel's decision-making reference.

[0040] It should be noted that the key technologies of this invention include: First, a physical-thermodynamic coupling model of the conductor state, which calculates the conductor temperature and sag correction by incorporating Joule heating, radiative heating, and wind speed heat dissipation into a unified heat balance equation, enabling the ranging benchmark to follow the changes in the physical state of the conductor in real time, thus eliminating the deviation between the static benchmark and the actual sag from the root cause; Second, a neural implicit network directed distance field reconstruction mechanism, which encodes local morphological features as contextual prior parameters and drives the neural implicit function to predict directed distance values, and combines curvature continuity constraints to fill in occlusion blind spots, so that the reconstructed surface remains geometrically continuous, avoiding the introduction of ranging deviations by holes; Third, a bidirectional feedback spatiotemporal deformation residual model, which couples the spatial geometry extraction branch and the time series long short-term memory branch in parallel and introduces the catenary physical regularization constraint, so that the network is simultaneously subject to the dual constraints of data-driven and physical laws, outputting a dynamic geometric benchmark with the ability to predict deformation trends in the time dimension. The three key technologies mentioned above work together to ensure the physical accuracy of the sag reference through thermodynamic correction, the geometric continuity of the surface through implicit reconstruction, and the predictive deformation of time series to ensure that the deformation trend is known under dynamic working conditions. Together, they support the adaptive spatial dynamic bounding box ranging algorithm driven by the electrical geometry model to continuously output reliable minimum spatial intervals in complex dynamic scenarios. The overall ranging accuracy is substantially improved compared with the single static reconstruction scheme.

[0041] It should be noted that under conditions where the dynamic range of conductor sag is large and the spatial distance from the intruder is near a critical value, the ranging system is extremely sensitive to the correction amount of conductor sag. Any deviation in sag estimation caused by temperature model errors or point cloud occlusion may lead to false alarms or missed alarms. The reason for the above technical problem is that the conductor sag fluctuates continuously with the coupled changes of current load, temperature, solar radiation intensity, and wind speed. Traditional point cloud reconstruction methods only capture the static geometric coordinates at the current moment and do not couple the thermodynamic physical state with the geometric reconstruction process. When the actual sag of the conductor increases significantly under heavy load or high temperature conditions, the deviation between the conductor position in the reconstructed model and the actual position can reach tens of centimeters or even more, causing a systematic shift in the ranging results within the safety critical range. The usual solution to the above technical problem is to perform geometric fitting on the conductor trajectory after point cloud acquisition and make simple linear corrections to the sag using empirical coefficients. However, empirical coefficients are difficult to cover multi-dimensional coupled scenarios with different loads, temperatures, and solar radiation. When the combination of operating conditions changes, the correction residuals are still significant and cannot meet the ranging accuracy requirements under critical interval scenarios. This invention effectively solves this technical problem by establishing a physical-thermodynamic coupling model of the conductor state. Joule heating, radiative heating, and wind speed heat dissipation are incorporated into a unified thermal balance equation for coupled solution. The temperature change of the conductor is calculated from the thermal balance result, and then converted into a sag correction amount from the material's thermal expansion and contraction coefficient. This ensures that the geometric position of the conductor in the reconstructed space always corresponds to the current physical state. At the same time, a two-way feedback spatiotemporal deformation residual model predicts the deformation trend from a time series dimension. The catenary physical equation is introduced as a regularization constraint, so that the deformation prediction is also constrained by physical laws. This further compresses the sag estimation residual under critical interval scenarios, thereby continuously ensuring the accuracy of the minimum spatial interval measurement under multi-dimensional working condition coupling.

[0042] 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 real-time three-dimensional reconstruction and ranging of potential hazards in power transmission channels based on fixed lidar.

[0043] A third aspect of the present invention provides a real-time three-dimensional reconstruction and ranging system for potential hazards in power transmission channels based on fixed lidar, 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.

[0044] Specifically, the principle of this invention is as follows: The fundamental reason why this invention can solve the above-mentioned technical problems lies in the synergistic operation of the physical thermodynamic coupling mechanism, the neural implicit geometric reconstruction mechanism, and the deep learning temporal deformation prediction mechanism within the same reconstruction framework. These three mechanisms jointly support the accuracy of the dynamic ranging benchmark from different dimensions. First, the fundamental reason for the deviation of the static ranging benchmark is that the reconstruction process only uses the geometric coordinates of the point cloud at the current moment, without physically modeling the sag change of the conductor due to thermal expansion. This invention establishes a physical thermodynamic coupling model of the conductor state, calculates the conductor temperature change based on the thermal balance relationship of Joule heating, radiative heating, and wind speed heat dissipation, and then converts it into a sag correction amount using the material's thermal expansion and contraction coefficient. This allows the geometric position of the conductor in the reconstruction space to be updated in real time with the environment and load state at any given moment, eliminating the influence of sag deviation on the ranging benchmark from a physical mechanism perspective. Second, point cloud occlusion blind spots cause surface discontinuities, further introducing ranging deviations. This invention regresses a directed distance field using a neural implicit network, encoding local morphological features as contextual prior parameters. This drives the network to predict the offset of the query point relative to the latent surface. Combined with curvature continuity constraints, matching and completion are performed in blind zones. Furthermore, a moving cube algorithm and non-uniform rational B-splines are used to mesh and fit the implicit field to the surface, outputting a geometrically continuous 3D surface model of the channel. This eliminates the interference of holes on ranging accuracy from a geometric reconstruction perspective. Secondly, conductor deformation exhibits inertia and trends over time, and point clouds at a single moment cannot reflect the deformation evolution. This invention uses a bidirectional feedback spatiotemporal deformation residual model, coupling the 3D spatial geometry extraction branch and the time series long short-term memory branch in parallel. The catenary physical equation is introduced as a regularization constraint, ensuring that the network output deformation trend is simultaneously constrained by both data-driven and physical laws, providing a time-dimensional deformation prediction basis for ranging. Finally, the adaptive spatial dynamic bounding box ranging algorithm driven by the electrical geometry model combines the electrical insulation safety distance standard with a hierarchical bounding box tree. It achieves efficient coarse screening by determining whether the bounding box of an intruder intersects with the bounding boxes of electrical wires and conduits, and then activates the local shortest path algorithm to finely solve for the minimum spatial interval, meeting real-time response requirements while ensuring accuracy. The above four mechanisms are logically progressive: thermodynamic correction ensures accurate sag reference, implicit reconstruction ensures surface continuity, temporal deformation prediction ensures the deformation trend is known in dynamic scenarios, and bounding box ranging ensures high efficiency and accuracy. The synergy of these four mechanisms enables the entire system to continuously output a reliable minimum spatial interval under complex dynamic conditions, thus meeting the inherent requirements for solving the aforementioned technical problems at both the logical and mechanistic levels.

[0045] 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.

[0046] The specific implementation of step S1 is as follows: a fixed lidar sensing unit is installed at the top of the transmission tower, a three-dimensional spatial coordinate system is established with the top of the tower as the origin, a panoramic scan of the transmission corridor area is performed, and an initial three-dimensional point cloud dataset is collected. In the formula For the first A four-dimensional attribute vector of points, For spatial coordinates, Reflection intensity, This represents the total number of point clouds.

[0047] The specific implementation of step S2 is to fuse local geometric feature entropy and density non-uniformity, and then... Perform adaptive segmentation. Local geometric feature entropy. The statistical dispersion of reflection intensity, normal vector angle, and curvature distribution within a local neighborhood of a point cloud is reflected by the following formula: ; In the formula, To count the number of boxes, For the first The point is in the local neighborhood Inner The normalized probabilities of each bin are statistically analyzed based on the distribution of reflection intensity, the distribution of the angle between normal vectors, and the curvature distribution. The average value is taken as the final local geometric feature entropy. A higher value indicates a more irregular geometric shape in the region. Density inhomogeneity The formula for calculating the density difference of point cloud distribution within a unit space is as follows: ; In the formula, For point In the local neighborhood Inner The point density of individual voxels, in units of , For the standard deviation operator, For the mean operator, Dimensionless. and The joint input consists of a classification framework based on region growing and graph cut theory. Region growing progressively expands the set of similar points based on neighborhood similarity, while graph cut transforms the segmentation problem into a minimum cut problem in graph theory. This decouples power grid equipment and facility points from vegetation points, outputting a set of vegetation points and a set of equipment and facility points. For redundant points in the equipment and facility point set, based on spatial grid partitioning and point distance analysis, points with Euclidean distances less than a preset threshold are... The points are merged. Filtering experiments were conducted on point clouds of multiple typical scenes, and the surface error before and after filtering was statistically analyzed. The empirical range was 0.01–0.05 m. An adaptive parametric density clustering algorithm was used for discrete noise points, accelerating neighborhood lookup through a spatial index structure and dynamically determining the cluster radius. With minimum number of points threshold Both were obtained through statistical analysis of point cloud data from multiple scenarios. A feature-preserving smoothing algorithm was used to smooth high-frequency disturbance noise based on local curvature. The normal vector is used to distinguish flat regions from feature edges, and a differentiated smoothing intensity factor is employed. Perform filtering. The calculation formula is as follows: ; In the formula, Based on the smoothness strength, dimensionless. This is the curvature sensitivity coefficient, in units of... (with curvature) unit (After multiplication, the exponent becomes dimensionless). Both were obtained through multiple filtering experiments and error statistical analysis. The empirical range is 0.3 to 0.8. The empirical range is 0.5–2.0 m, and the output is a clean point cloud dataset. .

[0048] The specific implementation of step S3 is as follows: The 3D trajectory of the traverse is obtained from the candidate point set of the traverse using the catenary physical constraint cutting plane reconstruction algorithm. The equation of the 3D nonlinear catenary is as follows: ; In the formula, This is the curve tensor parameter, in units of... , This is the horizontal offset, in units of , This is the vertical offset, in units of... , The coordinates are along the horizontal direction of the conductor, in units of , The coordinates are in the vertical direction, and the units are... The initial value range was obtained through multiple curve fitting experiments on historical traverse point cloud data combined with least squares error analysis, and then determined through iterative correction using field measurement data. Each iteration is based on candidate points. Euclidean distance to the currently fitted catenary trajectory (Unit is) Adaptively construct the convex hull cutting plane to handle distances exceeding a threshold. The impurities were cut and separated. Point clouds from multiple typical scenarios were obtained through experimental statistical analysis. The optimal catenary parameter vector was solved iteratively using the Newton-Raphson method. The iteration format is as follows: ; In the formula, For the first The residual vector of the nth iteration, in units of , This is a Jacobian matrix, with elements representing pairs of catenary equations. , , The partial derivatives of are dimensionless. [Number] iterations, iterating to... The process terminates when the convergence threshold is reached.

[0049] The specific implementation of step S4 is as follows: hierarchical feature learning is used to... Local morphological features are extracted to drive a hidden neural network to regress a directed distance field. This is then applied to spatial query points. Neural query networks rely on prior parameters from local context. ( Predict offset (for feature dimension) The query point after offset is ,Will Common input neural implicit function Predict the directed distance value , Units are The joint loss function is as follows: ; In the formula, For spatial projection loss, For distance on the real curved surface Closest point Units are For the directed distance field regression loss, This represents the true directed distance value, in units of , Units are For implicit surface constraint loss, the gradient magnitude of the directed range field must be approximately 1. Dimensionless , , Here are the harmonic weighting coefficients for each loss term, where , Units are , Dimensionless, obtained through multiple training experiments on the validation set and combined with convergence performance analysis, makes... The overall dimensionless property. The moving cube algorithm spatially divides the global directed distance field using a regular 3D mesh. It queries the sign and value of the directed distance field at the eight vertices of a voxel, determines the topology of triangular facet connections using a lookup table based on a preset isosurface threshold, and calculates the vertex coordinates of the facets using linear interpolation. A multi-resolution subdivision strategy is applied to the traverse region to generate a high-fidelity triangular mesh. Finally, non-uniform rational B-splines are used to adaptively fit and complete the blind zone, outputting a continuous channel 3D surface model. .

[0050] The specific implementation of step S5 is as follows: After identifying the intruder, a coupled physical-thermodynamic model of the conductor's state is established. Based on the principle of conductor thermal balance, the Joule heat generated by the current load, the heat absorbed by solar radiation, and the heat dissipation due to wind speed are incorporated into a unified thermal balance equation, which is as follows: ; In the formula, The mass of a conductor per unit length, in units of , This refers to the specific heat capacity of the conductor, in units of... , Temperature of the conductor, in units of , For time, the unit is , For current load, the unit is... , The resistance per unit length of wire, in units of . , The heat power absorbed by solar radiation per unit length, in units of , The convective heat dissipation power per unit length, in units of , Wind speed, unit: , Ambient temperature, unit: The coefficients were obtained through laboratory thermal experiments on the conductor material and regression analysis based on long-term field monitoring data. The dimensions on both sides of the equation are 0.5. Sag correction amount The calculation formula is as follows, based on the coefficient of thermal expansion and contraction of the conductor material: ; In the formula, The coefficient of linear thermal expansion of the conductor material, in units of 1000 kJ / m². , This refers to the distance between the gears, in units of... , For reference temperature, the unit is... , Units are , The distance was obtained through laboratory thermal experiments. In the adaptive spatial dynamic bounding box ranging algorithm driven by the electrical geometry model, the dynamic electrical insulation distance corresponding to the conductor voltage level is used. (Unit is) ) and wind deflection angle (Unit is) This forms an irregularly shaped spatial safety conduit enclosure surrounding the conductor axis. , Based on relevant safety regulations in the power industry and corrected using on-site meteorological experimental data, the intruder point cloud was spatially partitioned using a hierarchical bounding box tree to determine the intruder bounding boxes. and If the paths intersect, activate the local Dijkstra's shortest path graph theory algorithm if they intersect, and construct a weighted graph within the intersecting subspace. In the formula For a set of nodes, Let be the set of edges. This is the set of edge weights, where each edge weight is the Euclidean distance between the corresponding two nodes, in units of 1. The shortest dynamic three-dimensional spatial distance is solved using Dijkstra's algorithm. (Unit is) ), which serves as the minimum spatial interval between the intruder and the conductor.

[0051] In the bidirectional feedback spatiotemporal deformation residual model, the spatial geometry extraction branch is used to sequence temporal point clouds. ( (Based on the sampling time index) Features are extracted layer by layer. The time series long short-term memory branch combines timestamps, wind speed, and temperature time series data to model the deformation trend. Cross-layer jump connections and inverse correction loops enable the deformation acceleration trend output by the time series branch to reflect the deformation acceleration trend. (Unit is) The weights of the spatial features in the first few layers are fine-tuned using the gate signal, and the loop iterates 3 times. The loss function incorporates regularization constraints based on the catenary physical equation. The total loss is as follows: ; In the formula, To predict loss for deformation, This is the loss due to the deviation between the network output conductor shape and the theoretical shape of the catenary equation; both have the same dimensions. These are regularization weight coefficients, dimensionless, obtained through multiple training experiments on the validation set. Resource adjustment value. The calculation formula is as follows: ; In the formula, This represents the current batch of point cloud data, expressed in units of points. The number of iteration layers for the network residual blocks. The number of activations for cross-layer skip connections. , , For the dimensional harmonic weighting coefficients, make These are dimensionless numerical values. They were obtained through multiple training and inference experiments on point cloud data of different scales in an experimental environment, and by statistically analyzing the relationship between resource consumption and computation time. At that time, a single CUDA stream is used for serial execution. When dual CUDA streams are allocated in parallel, Multiple CUDA stream allocation is used.

[0052] The specific implementation of step S6 is as follows: based on the minimum spatial interval Based on the security situation level, the alarm system executes tiered warnings and generates handling suggestions. The security situation level is determined by spatiotemporal overlay analysis of the intruder's predicted trajectory and the conductor's spatial attitude, combined with the probability of spatial intersection. and Divided into discrete levels, The grading threshold was determined by statistically analyzing historical hazard cases and verifying them through multiple probabilistic simulations of the predicted trajectory and conductor attitude.

[0053] To better understand and implement this invention, the following is a specific application scenario of this invention, Example 2: With a certain 220 The transmission line corridor is the test scenario, and the distance between the towers within the corridor is approximately 450 mm. The conductor type is steel-cored aluminum stranded wire, and the conductor cross-sectional area is 300. The maximum sag in the design is approximately 8. The fixed lidar sensing unit is installed at the top of the tower in the middle of the corridor, with a scanning radius covering approximately 230 half-span. The sampling range was set at a point cloud sampling frequency of 10 times / second, with approximately 800,000 points scanned per scan. The test was conducted under high-temperature, high-load conditions during summer, with an ambient temperature of 38°C and a real-time conductor current load of 820kJ / m³. The solar radiation intensity is 950. The wind speed is 2.3. There is a tree in the passage whose canopy height encroaches on the clearance as an intrusion target.

[0054] After step S1, the initial 3D point cloud dataset acquired by the LiDAR includes mixed terrain features such as trees, power lines, poles, and the ground within the corridor, totaling approximately 800,000 points. After step S2, based on local geometric feature entropy and density non-uniformity, a joint classification framework of region growing and graph cutting is used to decouple vegetation points from power grid equipment and facility points. After three-level filtering, clean point cloud data is output, reducing the number of points to approximately 180,000. After step S3, the catenary physical constraint cut plane reconstruction algorithm converges after 5 iterations, with a cut plane threshold of 0.5. Newton-Raphson method to update catenary morphological parameters , , Afterwards, the 3D trajectory reconstruction of the conductor was completed, with an impurity point removal rate of over 97%. Following step S4, the neural implicit network continuously completed the blind spot on the back side of the tower, outputting a geometrically continuous 3D surface model of the channel. The weight ratio of the three terms in the joint loss function of the neural implicit network was 1:1:0.1. The comparison of surface continuity before and after blind spot completion is shown in Table 1.

[0055] Table 1 Comparison of Surface Continuity Before and After Blind Zone Completion

[0056] In step S5, the physical-thermodynamic coupling model of the conductor state is based on the current current 820. Ambient temperature 38℃, solar radiation intensity 950 Wind speed 2.3 By solving the thermal balance equations in a coupled manner, the conductor temperature rise was found to be approximately 62°C, with a corresponding sag correction of 0.43. This means that under the current operating conditions, the actual sag of the conductor is 0.43 mm more than the design value. The geometric position of the conductor in the reconstructed space is synchronously adjusted downwards. The bidirectional feedback spatiotemporal deformation residual model receives the past 30 frames (i.e., 3... The time-series point cloud coordinates and synchronous wind speed and temperature series data are used to output deformation trends after three iterations, predicting the next 5 days. The sway amplitude of the inner conductor under wind speed fluctuations is approximately 0.12. The conductor sag correction amounts under different operating conditions are shown in Table 2.

[0057] Table 2 Comparison of Conductor Sag Correction Amounts under Different Working Conditions

[0058] An adaptive spatial dynamic bounding box ranging algorithm driven by an electrical geometry model is based on 220 The dynamic electrical insulation distance corresponding to the voltage level and the current wind deflection angle are used to construct an irregular spatial safety conduit bounding box around the axis of the non-uniform rational B-spline curve of the conductor. The intruder point cloud is spatially partitioned using a hierarchical bounding box tree. After coarse screening to determine if the intruder bounding box intersects with the conductor / electrical conduit bounding box, a local shortest path graph theory algorithm is activated to finely solve for the minimum dynamic three-dimensional spatial distance. The comprehensive thermodynamic sag correction is 0.43. The deformation trend prediction swing is 0.12. Subsequently, the minimum spatial interval between the highest point of the intruding tree's crown and the guide wire was calculated to be 2.17. The next 5 The minimum interval for internal predictions may narrow to 2.05. All are below 220 The standard safety distance for the line is 4.0. .

[0059] like Figure 2 As shown, this is a comparative diagram of the sag correction for different operating conditions in this test scenario. The horizontal axis represents the operating condition number, and the vertical axis represents the sag correction. The diagram clearly shows the distribution of the sag correction output by the thermodynamic coupling model under different combinations of current, temperature, solar radiation, and wind speed.

[0060] like Figure 3 As shown, the curves of deformation prediction residual convergence of the bidirectional feedback spatiotemporal deformation residual model in the three iterations under this test scenario are shown. The horizontal axis is the time frame number and the vertical axis is the deformation prediction residual. The figure shows the gradual convergence process of the residuals in the three iterations, which verifies the gradual correction effect of the bidirectional feedback mechanism on the deformation reconstruction accuracy.

[0061] After step S6 is completed, based on the minimum spatial interval of 2.17... Based on the threshold for classifying security status, if the security status is determined to be Level 3 (high risk), the alarm system will output a Level 3 warning and generate a handling suggestion, prompting maintenance personnel to carry out emergency pruning of the target trees and increase the frequency of real-time monitoring.

[0062] Compared to traditional static point cloud ranging methods, this invention embeds a physical-thermodynamic coupling mechanism into the 3D reconstruction process, enabling the ranging benchmark to follow the real physical changes in the conductor sag in real time. This eliminates the systematic sag estimation bias caused by neglecting thermodynamic effects at the fundamental level. By using a neural implicit network for directed distance field reconstruction and curvature continuity constraints to fill blind spots, it avoids the disruption of surface geometric continuity caused by occlusion holes, ensuring the integrity of the ranging benchmark from a geometric perspective. Furthermore, by introducing a two-way feedback spatiotemporal deformation residual model to predict deformation trends in the time dimension, the system possesses the ability to dynamically perceive the future spatial position of the conductor, a capability that traditional static methods cannot provide. These three mechanisms logically compensate for the inherent deficiencies of static ranging methods in the physical, geometric, and temporal dimensions, ensuring that the measurement results of the minimum spatial interval remain consistent with the actual spatial position of the conductor under complex dynamic conditions. This achieves highly reliable real-time 3D reconstruction and ranging of potential hazards in power transmission channels.

[0063] It should be noted that the variables involved in this invention are explained in detail in Tables 2 and 3.

[0064] Table 2. Variable Explanation Table (Part 1)

[0065] Table 3. Variable Explanation Table (Part Two)

[0066] 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 real-time three-dimensional reconstruction and ranging of potential hazards in power transmission channels based on fixed lidar, characterized in that, Includes the following steps: A fixed lidar sensing unit is installed on the top of the transmission tower to perform a panoramic scan of the transmission corridor area and collect an initial three-dimensional point cloud dataset. By integrating local geometric feature entropy and density non-uniformity, the initial 3D point cloud dataset is adaptively segmented, separating vegetation points and power grid equipment and facility points, and then filtering out redundant points, discrete noise points and high-frequency disturbance noise in sequence to output clean point cloud data. For the candidate point set of the conductor in the clean point cloud data, the catenary physical constraint cutting plane reconstruction algorithm is used to remove impurity points that deviate from the physical trajectory, and the three-dimensional trajectory of the conductor is obtained. By capturing local morphological features of clean point cloud data through hierarchical feature learning, a neural implicit network is driven to regress the directed distance field. Blind zone matching and completion are performed by combining curvature continuity constraints. The moving cube algorithm and non-uniform rational B-splines are used to mesh and fit the implicit geometric expression to the surface, and a continuous channel three-dimensional surface model is output. Intruders or abnormal objects are identified in the three-dimensional surface model of the channel. A physical-thermodynamic coupling model of the conductor state is established. The theoretical sag correction is calculated based on the current load, ambient temperature, solar radiation intensity and wind speed. An adaptive spatial dynamic bounding box ranging algorithm driven by the electrical geometry model is used. Combined with the deformation trend output by the bidirectional feedback spatiotemporal deformation residual model, the minimum spatial interval between the intruder and the conductor is measured. Based on the minimum spatial interval and security status level, the alarm system is driven to issue graded warnings and generate handling suggestions.

2. The method for real-time three-dimensional reconstruction and ranging of hidden dangers in power transmission channels based on fixed lidar as described in claim 1, characterized in that, The local geometric feature entropy refers to the statistical dispersion of the reflection intensity, normal vector angle, and curvature distribution within a local neighborhood of a point cloud; the density non-uniformity refers to the statistical quantity of the difference in density of point cloud distribution per unit space.

3. The method for real-time three-dimensional reconstruction and ranging of hidden dangers in power transmission channels based on fixed lidar as described in claim 2, is characterized in that, The adaptive segmentation specifically involves inputting the local geometric feature entropy and density non-uniformity into a classification framework composed of region growing and graph cut theory. Region growing gradually expands the set of similar points based on neighborhood similarity, while graph cut transforms the segmentation problem into a minimum cut problem in graph theory. The combination of the two decouples the power grid equipment and facility points from the vegetation points.

4. The method for real-time three-dimensional reconstruction and ranging of hidden dangers in power transmission channels based on fixed lidar as described in claim 3, is characterized in that, The sequential filtering of redundant points, discrete noise points, and high-frequency disturbance noise specifically involves: filtering redundant points using spatial grid partitioning and point distance analysis, fusing points that fall within the same micro-element grid or whose Euclidean distance to neighboring points is less than a distance threshold; filtering discrete noise points using an adaptive parameter-based density clustering algorithm, accelerating neighborhood queries through a spatial index structure, and dynamically determining the clustering radius and minimum point threshold based on the point cloud distribution characteristics; and filtering high-frequency disturbance noise using a feature-preserving smoothing algorithm, distinguishing flat regions from feature edges based on local curvature and normal vectors, and using a differentiated smoothing intensity factor for filtering.

5. The method for real-time three-dimensional reconstruction and ranging of hidden dangers in power transmission channels based on fixed lidar according to claim 4, characterized in that, The physical-thermodynamic coupling model of the conductor state is established by incorporating the Joule heat generated by the current load, ambient temperature, solar radiation absorption heat and wind speed heat dissipation into a unified thermal balance equation for coupling and solving, thereby obtaining the conductor temperature change, and then converting the sag correction amount based on the thermal expansion and contraction coefficient of the conductor material.

6. The method for real-time three-dimensional reconstruction and ranging of hidden dangers in power transmission channels based on fixed lidar as described in claim 5, is characterized in that, The neural implicit network regresses the directed distance field by encoding local morphological features into local context prior parameters. The neural query network predicts the offset of the spatial query point relative to the potential surface based on the local context prior parameters. The offset query point position and the local context prior parameters are then input into the neural implicit function to jointly predict the directed distance value of the point.

7. The method for real-time three-dimensional reconstruction and ranging of hidden dangers in power transmission channels based on fixed lidar as described in claim 6, characterized in that, The neural query network and the neural implicit function are trained end-to-end through a joint loss function. The joint loss function consists of a weighted sum of spatial projection loss, directed distance field regression loss, and implicit surface constraint loss. The spatial projection loss calculates the square of the Euclidean distance between the offset query point and the nearest point on the true surface. The directed distance field regression loss calculates the square of the difference between the predicted directed distance value and the true directed distance value. The implicit surface constraint loss requires that the gradient magnitude of the directed distance field at the query point be close to 1.

8. The method for real-time three-dimensional reconstruction and ranging of hidden dangers in power transmission channels based on fixed lidar as described in claim 7, characterized in that, The moving cube algorithm specifically involves dividing the global directed distance field into a regular three-dimensional grid, querying the directed distance field sign and value at the eight vertices of the voxel, determining the topology of the triangular facets through a lookup table based on the isosurface threshold, and calculating the vertex coordinates of the facets through linear interpolation. For traverse regions with gentle gradient changes, a multi-resolution subdivision strategy is used to generate a high-fidelity triangular grid.

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 method for real-time three-dimensional reconstruction and ranging of power transmission channel hazards based on fixed lidar as described in any one of claims 1-8.

10. A real-time three-dimensional reconstruction and ranging system for potential hazards in power transmission channels based on fixed lidar, 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