A power transmission line working condition prediction analysis and risk level evaluation method, medium and system

Abstract

The application provides a power transmission line working condition prediction analysis and risk level evaluation method, medium and system, and belongs to the technical field of power transmission lines. The application collects three-dimensional point clouds, weather, load, temperature and historical inspection data of the power transmission line corridor, and completes space-time alignment. After multi-scale processing of the point clouds, multi-view pseudo images are generated. Object-level three-dimensional entities are obtained by inputting the atlas-constrained multi-view pseudo image detection network. A three-dimensional spatial relationship atlas with objects as nodes is constructed. The catenary-wind load coupling clearance prediction algorithm is called to calculate the conductor clearance, sag and wind deflection working condition characteristics and prediction interval. The object relationship graph shortest risk path propagation algorithm is called to calculate the multi-factor risk index and risk level. Finally, the risk evolution result and disposal suggestion are output through object identity inheritance, solving the technical problem that the accuracy and stability of power transmission line clearance prediction significantly decrease under data missing or shielding working conditions due to insufficient fusion of multi-source heterogeneous data.

Description

Technical Field

[0001] This invention belongs to the field of power transmission line technology, and specifically relates to a method, medium, and system for predicting and analyzing the operating conditions and assessing the risk level of power transmission lines. Background Technology

[0002] Predicting operating conditions and assessing risks in transmission line corridors are crucial aspects of power grid operation and maintenance. Existing technologies typically employ single-sensor meteorological or point cloud data collection, combined with a simplified catenary model to calculate conductor sag and clearance, and then classify risks based on fixed thresholds. Some solutions introduce 2D image detection or single-view point cloud segmentation to identify objects such as towers, conductors, and tree obstructions, and then combine this with meteorological load data to estimate clearance margin. However, due to the complex environment of transmission line corridors, including tree obstruction, multiple intersections, and sensor scanning blind spots, single-view detection may miss objects or have positioning errors in obstructed areas. Furthermore, meteorological and load data may be missing under extreme conditions, leading to increased catenary fitting residuals, uncontrolled clearance prediction intervals, and inaccurate tracing of risk propagation paths to the actual risk sources. In other words, existing technologies suffer from insufficient fusion of multi-source heterogeneous data, resulting in a significant decrease in the accuracy and stability of transmission line clearance prediction under data gaps or obstruction conditions. Summary of the Invention

[0003] In view of this, the present invention provides a method, medium and system for predicting and analyzing the operating conditions and assessing the risk level of transmission lines, which can solve the technical problem in the prior art where insufficient fusion of multi-source heterogeneous data leads to a significant decrease in the accuracy and stability of transmission line clearance prediction under conditions of missing data or obstruction.

[0004] The present invention is implemented as follows: The first aspect of the present invention provides a method for predicting and analyzing the operating conditions and assessing the risk level of transmission lines, comprising the following steps: Collect 3D point cloud data, meteorological data, load data, temperature data and historical inspection data of the transmission line corridor, perform time alignment and coordinate benchmark unification, and generate data quality labels; The 3D point cloud data is processed by denoising, downsampling, block indexing and multi-scale spatial representation, and a high point density is retained for conductor and insulator regions. Generate multi-view pseudo-images, input a graph-constrained multi-view pseudo-image detection network, and output object-level 3D entity representations of towers, conductors, insulators, tree barriers, and crossings; Point-level segmentation and boundary refinement are performed within the neighborhood of object-level 3D entities to construct a 3D spatial relationship graph with objects as nodes and spatial relationships as edges; Based on the three-dimensional spatial relationship map, the catenary-wind load coupled clearance prediction algorithm is called to calculate the characteristics of conductor clearance, sag, wind deflection and icing load conditions, as well as future predicted values ​​and confidence intervals. Based on the predicted working condition characteristics, measurement quality score and safety threshold, the object relationship graph shortest risk path propagation algorithm is invoked to calculate and output the multi-factor risk index and risk level. Output risk evolution results and disposal suggestions through object identity inheritance.

[0005] Specifically, the input of the graph-constrained multi-view pseudo-image detection network receives the forward line view, the transverse line view, and the top view. Figure 3 The pseudo-image group encodes five channels for each pixel: height, depth, reflection intensity, point density, and local height difference.

[0006] Specifically, the network front-end of the graph-constrained multi-view pseudo-image detection network consists of a lightweight convolutional backbone with four downsampling stages. Each stage is composed of depth-separable convolutional layers and batch normalization layers. An orientation-aware attention module is embedded in the middle to assign different channel weights and spatial weights to conductor features, insulator string features, and tower features, respectively.

[0007] Specifically, the detection head of the graph-constrained multi-view pseudo-image detection network adopts a decoupled structure, outputting two-dimensional candidate box coordinates, category confidence, and key point coordinates within the view respectively. After the detection head, a graph consistency constraint layer is set to perform soft matching on the candidate boxes output by the three views according to projection geometric constraints, category consistency, and topological adjacency relationships to generate cross-view association scores.

[0008] The graph-constrained multi-view pseudo-image detection network is equipped with an iterative fusion unit. The iterative fusion unit iterates for a preset number of rounds. In each round, if the confidence of a candidate box in a certain view is lower than the low confidence threshold, then the candidate regions in the other two views with confidence higher than the high confidence threshold are selected. An attention mask is generated by coordinate back projection and superimposed on the feature map corresponding to the low confidence view. The features of the view are recalculated and the candidate box and confidence are re-output.

[0009] Specifically, the resource allocation method of the iterative fusion unit is as follows: after each iteration, the video memory occupied by the intermediate feature map generated in that iteration is released, and the video memory allocation ratio of the released video memory is allocated to the CUDA stream corresponding to the attention mask calculation in the next iteration; the memory occupied by the view branch feature map that does not participate in the iterative fusion is marked as read-only during the iteration and its memory bandwidth allocation priority is reduced.

[0010] The training of the graph-constrained multi-view pseudo-image detection network adopts a phased training method: the first phase fixes the parameters of the graph consistency constraint layer and the iterative fusion unit, and only trains the convolutional backbone, attention module and detection head; the second phase unfreezes the graph consistency constraint layer and uses cross-view association loss in the form of contrast loss for joint training; the third phase unfreezes the iterative fusion unit and introduces low-confidence samples to train the network's ability to correct the features of low-confidence views using high-confidence views.

[0011] Specifically, the catenary-wind load coupled clearance prediction algorithm uses the coordinates of the hanging point, the coordinates of the lowest point, and the span length extracted from the point cloud to fit the catenary model. The fitting residual is used to generate the basic reliability. The thermal elongation correction is calculated based on the conductor temperature data, the equivalent vertical load correction is calculated based on the line load current and conductor material parameters, and the lateral wind displacement is calculated based on the wind speed and direction data. The three types of corrections are superimposed on the catenary baseline curve to obtain the corrected conductor prediction curve.

[0012] Specifically, when meteorological or load data is missing, the catenary-wind load coupled headroom prediction algorithm retrieves similar operating condition samples with similar temperature, load, and wind speed ranges from the historical database, takes the weighted average of the correction amount as compensation, and distributes the weights according to the similarity from high to low. The similarity threshold is determined by cluster analysis of the historical samples. The prediction interval width is determined by a combination of the basic confidence level and the number of compensation samples.

[0013] Specifically, the object relationship graph shortest risk path propagation algorithm uses tower nodes, conductor segment nodes, insulator nodes, tree cluster nodes, ground partition nodes, and crossing node as graph nodes, and uses clearance relationship, crossing relationship, splicing relationship, proximity relationship, and occlusion relationship as graph edges. Each edge is assigned four weights: safety margin, distance change rate, spatial sensitivity, and measurement quality.

[0014] Specifically, the propagation process of the object relationship graph shortest risk path propagation algorithm is as follows: starting from the risk source edge, it propagates along the graph edge to the adjacent node. During the propagation process, the risk value decreases step by step according to the attenuation coefficient, which is determined by the topological distance and spatial sensitivity between adjacent edges. If the cumulative risk value of an adjacent edge exceeds the threshold corresponding to its safety margin, the edge is included in the risk propagation path until the cumulative risk value is lower than the cutoff threshold or reaches the graph boundary node.

[0015] Specifically, the path sorting of the object relationship graph shortest risk path propagation algorithm involves recording all propagation paths, sorting the paths according to their cumulative risk weights, and determining the source and terminal objects corresponding to the path with the highest cumulative risk weight as the risk source and the affected objects, respectively. The priority of each node on the path is determined according to its weight.

[0016] The graph edge types of the three-dimensional spatial relationship graph include clearance relationship, crossing relationship, hanging relationship, proximity relationship and occlusion relationship. The graph is generated after point-level segmentation and boundary refinement are completed in the neighborhood of the object-level three-dimensional entity during graph construction.

[0017] Specifically, the establishment of the training dataset involves collecting 3D laser point cloud data of multiple power transmission line corridors, generating multi-view pseudo-images, and then manually annotating the 2D bounding boxes and categories. Cross-view association is performed on the bounding boxes of the same physical object in the three views, and the corresponding relationship is recorded as the supervision label of the map consistency constraint layer. Low-confidence samples are generated by artificially occluding some views of some objects. The training set and the validation set are divided according to the training and validation ratio.

[0018] The low confidence threshold is 0.5, the high confidence threshold is 0.7, the preset number of rounds is 2 to 3, the memory allocation ratio is 80%, the residual quantile threshold for basic confidence is the 80th quantile, and the training-validation split ratio is 8:2.

[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 predicting and analyzing the operating conditions of a power transmission line and assessing its risk level.

[0020] A third aspect of the present invention provides a transmission line operating condition prediction analysis and risk level assessment system, 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 graph-constrained multi-view pseudo-image detection network, utilizing complementary information between the forward, transverse, and top-view views. It then uses an iterative fusion unit to correct low-confidence detection results for occluded views, ensuring stable object-level 3D entity localization even under occlusion conditions. Furthermore, a catenary-wind load coupled clearance prediction algorithm introduces a historical similarity compensation mechanism. When meteorological or load data is missing, weighted historical samples replace the missing input, outputting a clearance prediction interval with uncertainty. The object relationship graph shortest risk path propagation algorithm further expands the risk index of isolated objects into risk chains within the object relationship network, enabling quantifiable tracing of risk sources and propagation paths. These three mechanisms work together to eliminate the interference of data loss and occlusion on the accuracy of results at each stage, from object detection and condition prediction to risk propagation. In summary, this invention solves the technical problem mentioned in the background art where insufficient fusion of multi-source heterogeneous data leads to a significant decrease in the accuracy and stability of transmission line clearance prediction under data loss or occlusion conditions. Attached Figure Description

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

[0023] Figure 2 This is a projection distribution diagram of pseudo-images for testing the three-view format.

[0024] Figure 3 A visualization of risk propagation paths in a three-dimensional spatial relationship map. 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 predicting and analyzing the operating conditions and assessing the risk level of transmission lines, provided by the first aspect of this invention. This method includes the following steps: S1. Collect 3D point cloud data, meteorological data, load data, temperature data and historical inspection data of the transmission line corridor, perform time alignment and coordinate benchmark unification, and generate data quality labels. S2. Perform noise reduction, downsampling, block indexing and multi-scale spatial representation processing on the three-dimensional point cloud data, and retain a high point density for the conductor and insulator regions; S3. Generate multi-view pseudo-images, input the graph-constrained multi-view pseudo-image detection network, and output object-level 3D entity representations of towers, conductors, insulators, tree barriers, and crossings. S4. Perform point-level segmentation and boundary refinement within the neighborhood of the object-level 3D entity to construct a 3D spatial relationship graph with objects as nodes and spatial relationships as edges. S5. Based on the three-dimensional spatial relationship map, call the catenary-wind load coupled clearance prediction algorithm to calculate the characteristics of conductor clearance, sag, wind deflection and icing load conditions, as well as future predicted values ​​and confidence intervals. S6. Based on the predicted working condition characteristics, measurement quality score and safety threshold, call the object relationship graph shortest risk path propagation algorithm to calculate the multi-factor risk index and risk level and output them.

[0027] Optionally, after outputting the risk level, the steps may also include: outputting the risk evolution results and handling recommendations through object identity inheritance.

[0028] The graph-constrained multi-view pseudo-image detection network is a deep learning-based object detection model used to fuse two-dimensional detection results from multi-view pseudo-images into object-level three-dimensional entities based on spatial relationship constraints. Its specific structure is as follows: the input receives the forward line view, the horizontal line view, and the top view. Figure 3The network uses a set of pseudo-images, with each pixel channel encoding five channels: height, depth, reflection intensity, point density, and local height difference. The network front-end is a lightweight convolutional backbone containing four downsampling stages, each consisting of depthwise separable convolutions and batch normalization layers, used to extract multi-scale texture features. A direction-aware attention module is embedded in the middle, assigning different channel and spatial weights to laterally extending wire features, vertically arranged insulator string features, and frame-like tower features, respectively. The detection head employs a decoupled structure, outputting two-dimensional candidate box coordinates, class confidence scores, and keypoint coordinates within the view. After the probe, a map consistency constraint layer is set up to perform soft matching on the candidate boxes output from the three views according to projection geometry constraints, category consistency, and topological adjacency, generating cross-view association scores. An iterative fusion unit is set up within the network, iterating 2-3 times. In each round, if the confidence of a candidate box in one view is lower than 0.5, candidate regions with confidence higher than 0.7 from the other two views are selected, and attention masks are generated through coordinate backprojection. These masks are then superimposed on the feature map corresponding to the low-confidence view, and the features of that view are recalculated, and candidate boxes and confidence scores are re-output. The final output includes the object category, 3D candidate space range, and view... Figure 1 Consistency scoring. In terms of resource allocation, the number of iteration rounds of the iterative fusion unit is functionally related to the memory allocation. After each iteration, the memory occupied by the intermediate feature maps generated in that round is released, and 80% of the released memory is allocated to the CUDA stream corresponding to the next round of attention mask calculation. The memory occupied by the feature maps of the view branches that do not participate in the iterative fusion is marked as read-only during the iteration and its memory bandwidth allocation priority is reduced, thereby concentrating the computing resources on the correction calculation of low confidence views.

[0029] The steps for establishing the training dataset for the graph-constrained multi-view pseudo-image detection network specifically include: collecting 3D laser point cloud data of multiple transmission line corridors and generating multi-view pseudo-images as described in step S3; having manual annotators label each pseudo-image with 2D bounding boxes and categories of towers, conductors, insulators, fittings, tree barriers, and crossings; manually associating the bounding boxes of the same physical object across the three views and recording their correspondence as supervision labels for the graph consistency constraint layer; artificially occluding parts of the views for some objects to generate low-confidence samples for training iterative fusion units; and dividing the labeled dataset into training and validation sets in an 8:2 ratio.

[0030] The training steps of the graph-constrained multi-view pseudo-image detection network specifically include: a phased training approach; in the first phase, fixing the parameters of the graph consistency constraint layer and the iterative fusion unit, training only the convolutional backbone, attention module, and detection head to enable the network to have basic detection capabilities; in the second phase, unfreezing the graph consistency constraint layer, training the network by combining the detection loss and cross-view association loss, with the cross-view association loss adopting a contrastive loss form to bring the candidate box features of the same object closer in different views and push the feature distance of different objects further apart; in the third phase, unfreezing the iterative fusion unit, introducing low-confidence samples, and training the network's ability to correct low-confidence view features using high-confidence views; during training, a learning rate decay strategy is adopted, and the final model parameters are determined by using detection precision, cross-view association accuracy, and small object recall as comprehensive indicators on the validation set.

[0031] The technical benefits of the graph-constrained multi-view pseudo-image detection network are as follows: by enhancing the feature representation of slender and small-scale objects such as wires and insulators in pseudo-images through the direction-aware attention module, it alleviates the problem of small target features being submerged after point cloud projection; through the graph consistency constraint layer and iterative fusion unit, it uses complementary information between multiple views to correct view results with local occlusion or insufficient detection confidence, making the object-level 3D entity localization more stable, and providing a more reliable geometric input basis for subsequent spatial relationship graph construction and working condition prediction.

[0032] The principle of the catenary-wind load coupled clearance prediction algorithm is that the shape of the conductor within the span is mainly determined by its own weight load, and its static shape can be described by the catenary model, which serves as a reference curve. When the conductor temperature rises, the conductor material undergoes thermal elongation, increasing the sag. When wind load is present, the vertical load and horizontal wind pressure in the vertical plane work together to cause the conductor to deviate in the composite direction. This deviation is obtained by superimposing the equivalent vertical load component and the lateral wind deflection component. The algorithm first uses the coordinates of the hanging point, the coordinates of the lowest point, and the span length extracted from the point cloud to fit the catenary model. The fitting residual is used to generate the basic reliability. The smaller the residual, the higher the basic reliability. This threshold is determined by statistically analyzing the residual distribution and taking the 80th percentile of the point cloud fitting experiments of multiple conductors with known shapes. Subsequently, the algorithm calculates the thermal elongation correction based on conductor temperature data, the equivalent vertical load correction based on line load current and conductor material parameters, and the lateral wind displacement based on wind speed and direction data. These three corrections are then superimposed onto the catenary baseline curve to obtain the corrected conductor prediction curve. The algorithm performs a minimum distance search along the entire length of the prediction curve with ground point clouds, tree obstruction point clouds, building point clouds, and intersection point clouds to obtain the predicted values ​​of ground clearance and lateral clearance. When meteorological or load data is missing, the algorithm retrieves similar operating condition samples with similar temperature, load, and wind speed ranges from the historical database, and uses the weighted average of their corrections as compensation. The weights are distributed according to similarity from high to low, and the similarity threshold is determined by cluster analysis of historical samples, using the lower limit of the average similarity of samples within the same cluster. The algorithm finally outputs the predicted clearance value and its prediction interval. The width of the prediction interval is determined by a combination of the basic confidence level and the number of compensation samples; the lower the basic confidence level or the fewer the compensation samples, the larger the interval width. This algorithm combines the physical morphology of conductors with meteorological load factors, making the airspace prediction results physically interpretable. Furthermore, through a historical sample compensation mechanism, it can still output prediction results with uncertainty even when data is missing, thereby improving the stability and applicability of the prediction.

[0033] The principle of the object relationship graph shortest risk path propagation algorithm is that the risks between objects in a transmission line are not isolated. An anomaly in a certain relationship edge will be propagated to adjacent objects through spatial adjacency, topological connection, and other relationships. The algorithm uses tower nodes, conductor segment nodes, insulator nodes, tree cluster nodes, ground partition nodes, and crossing nodes as graph nodes, and clearance relationships, crossing relationships, splicing relationships, proximity relationships, and shading relationships as graph edges. Each edge is assigned four weights: safety margin, distance change rate, spatial sensitivity, and measurement quality. The algorithm first identifies edges with risk indices higher than a preset level as risk source edges. Starting from the risk source edge, the risk propagates along the graph edges to adjacent nodes. During the propagation process, the risk value decreases step by step according to the attenuation coefficient. The attenuation coefficient is determined by the topological distance and spatial sensitivity between adjacent edges. The greater the topological distance or the lower the spatial sensitivity, the greater the attenuation. The range of values ​​for the attenuation coefficient is determined by backtracking the propagation paths of historical risk events and statistically analyzing the degree of risk correlation under different topological distances. During the propagation process, if the cumulative risk value of an adjacent edge exceeds its own safety margin threshold, that edge is included in the risk propagation path and continues to propagate to the next level node until the cumulative risk value falls below the cutoff threshold or reaches the graph boundary node. The algorithm records all propagation paths and sorts them according to their cumulative risk weights. The source and terminal objects corresponding to the path with the highest cumulative risk weight are identified as the risk source and affected objects, respectively. The priority of handling each node on the path is determined according to its weight. This algorithm extends the risk of a single relational edge to a risk chain in an object relationship network, enabling the risk assessment results to reflect the risk source, propagation path, and scope of collateral impact, providing a basis for prioritization in operation and maintenance.

[0034] The specific implementation of step S1 is as follows: Three-dimensional point cloud data is collected along the transmission line corridor using an airborne lidar. Simultaneously, wind speed, wind direction, and temperature data from meteorological stations are recorded. Line load current data is obtained from the dispatch system, and historical inspection records for the same line are extracted from the historical inspection database. After data collection, interpolation and alignment are performed based on the timestamps of each data source. The point cloud coordinates are converted to meteorological station coordinates using the geodetic coordinate system as a unified coordinate reference to ensure spatial consistency. A data quality identifier is generated for each data record. The identifier includes sensor operating status, signal strength, missing measurement flags, and coordinate transformation residuals, providing a priori credibility of the input data for subsequent algorithms.

[0035] The specific implementation of step S2 is as follows: First, statistical filtering is used to remove outlier noise points from the original 3D point cloud data. The number of neighboring points within the neighborhood radius of each point is counted, and points with a neighboring point count lower than the lower threshold of point density are marked as noise and removed. Then, a voxel grid downsampling method is used to uniformly thin the point cloud. For the semantic regions of conductors and insulators that have been manually defined a priori, the voxel size is appropriately reduced to retain a higher local point density. After downsampling, the point cloud is divided into blocks and indexed according to an octree structure. Each leaf node corresponds to a point set with a fixed spatial range for subsequent fast neighborhood retrieval. Multi-scale spatial representation is achieved by extracting local geometric features at different voxel resolutions. Each point is accompanied by multi-scale descriptors such as height value, normal vector, and local height difference for subsequent pseudo-image generation.

[0036] The specific implementation of step S3 is as follows: the preprocessed point cloud is orthogonally projected onto three projection planes: along the route, perpendicular to the route, and vertically, to generate a view along the route, a view across the route, and a top view. Figure 3 A set of pseudo-images is generated, where each pixel of each pseudo-image is encoded with the mean height, mean depth, mean reflection intensity, point density, and local height difference of the point set within the spatial grid, forming a five-channel input tensor. These three sets of pseudo-images are input into a map-constrained multi-view pseudo-image detection network. The network extracts multi-scale texture features through a lightweight convolutional backbone. A direction-aware attention module assigns higher horizontal channel weights to laterally extending wire features and higher vertical spatial weights to vertically arranged insulator string features. The detection head outputs two-dimensional candidate boxes and class confidence scores for each view in a decoupled structure. A map consistency constraint layer performs soft matching on the candidate boxes of the three views based on projection geometry constraints. For low-confidence views, an iterative fusion unit is initiated, back-projecting the candidate regions of high-confidence views as attention masks, superimposing them onto the feature maps of low-confidence views, and then re-inferring. This process iterates for a maximum of three rounds, ultimately outputting the three-dimensional candidate spatial range and viewport of each physical object. Figure 1 Consistency scoring is used to form an object-level three-dimensional entity representation.

[0037] The specific implementation of step S4 is as follows: Within the spatial range of each object-level 3D entity, the original point cloud of the neighborhood is obtained using the nearest neighbor search method. The boundary of the object is refined using the region growing segmentation method on the neighborhood point cloud. The segmentation criteria are based on normal vector similarity and curvature continuity, and the boundary points are refined to centimeter-level accuracy. After object recognition and boundary refinement are completed, based on the spatial position and geometric features of each object, the following relationships are determined: clearance relationship (distance between the conductor and the ground, the conductor and the tree barrier), crossing relationship (positional relationship between the conductor and the crossing object), connection relationship (connection relationship between the conductor and the insulator, the insulator and the tower), proximity relationship (spatial adjacency relationship between adjacent objects), and occlusion relationship (object pairs that are occluded in the point cloud acquisition). Using the object as a node and the above five types of spatial relationships as directed edges, a 3D spatial relationship graph is constructed. The graph is stored as an adjacency matrix and an edge attribute dictionary to support subsequent graph algorithm calls.

[0038] The specific implementation of step S5 is as follows: The catenary-wind load coupled clearance prediction algorithm takes the coordinates of the hanging points, the coordinates of the lowest point, and the span length of the conductor segment nodes in the three-dimensional spatial relationship map as input. It uses the least squares method to fit the catenary model parameters, and the fitting residual is used as the basis for basic reliability. The smaller the residual, the higher the basic reliability. The basic reliability threshold is taken as the 80th quantile of the fitted residual distribution. After determining the catenary baseline curve, the thermal elongation correction is calculated based on temperature data, the equivalent vertical load correction is calculated based on load current and conductor material parameters, and the lateral wind displacement is calculated based on wind speed and direction data. These three types of corrections are superimposed on the baseline curve to obtain the prediction curve. The algorithm performs a minimum distance search along the entire length of the prediction curve with the ground point cloud, tree obstacle point cloud, building point cloud, and intersection point cloud, outputting the predicted values ​​and prediction intervals for ground clearance and lateral clearance. The interval width increases as the basic reliability decreases or the number of compensation samples decreases. When meteorological or load data is missing, samples with similar operating conditions in similar temperature, load, and wind speed ranges are retrieved from the historical database. The weighted average of their correction values ​​is used as compensation. The weights are distributed from high to low similarity. The similarity threshold is determined by taking the lower limit of the average similarity of samples within the same cluster after performing cluster analysis on the historical samples.

[0039] The specific implementation of step S6 is as follows: The object relationship graph shortest risk path propagation algorithm takes a three-dimensional spatial relationship graph as input. Each graph edge is assigned four weights: safety margin, distance change rate, spatial sensitivity, and measurement quality. The algorithm comprehensively calculates the risk index of each edge. The algorithm first identifies edges with risk indices higher than a preset risk level threshold as risk source edges. Starting from the risk source edge, the risk propagates along the graph edge to adjacent nodes. During propagation, the risk value decreases step by step according to the attenuation coefficient, which is determined by both topological distance and spatial sensitivity. The value of the attenuation coefficient is determined through backtracking analysis of historical risk event propagation paths. If the cumulative risk value of an adjacent edge exceeds its corresponding safety margin threshold, the edge is included in the risk propagation path. Propagation continues until the cumulative risk value is lower than the cutoff threshold or reaches a graph boundary node. The algorithm records all propagation paths, sorts them according to cumulative risk weights, and marks the source object and terminal object of the path with the highest weight as the risk source and the affected object, respectively. The nodes on the path are prioritized according to their weights, and finally, the multi-factor risk index and risk level are output. Subsequently, through the object identity inheritance mechanism, the risk level changes of each object in each assessment are recorded as a risk evolution sequence. Based on the risk level and object type, corresponding disposal suggestions are matched from the disposal rule base and output to the operation and maintenance personnel.

[0040] It should be noted that the key technologies of this invention include: the graph-constrained multi-view pseudo-image detection network, through a direction-aware attention module and an iterative fusion unit, transforms the complementary information of multi-view projections into detection correction signals for low-confidence views, enabling the 3D localization of objects under occlusion conditions to shift from relying on the integrity of a single view to relying on the geometric consistency of multiple views, fundamentally severing the direct impact of occlusion on detection stability; the catenary-wind load coupled clearance prediction algorithm embeds physical mechanisms into the clearance calculation framework, so that compensation when meteorological load is missing no longer relies on statistical extrapolation but on physical similarity, and the correction amount of historical similar condition samples has bounded error under the constraints of physical laws, and the expansion direction of the prediction interval is consistent with the direction of uncertainty of the missing data, ensuring the physical interpretability of the prediction results; the object relationship graph shortest risk path propagation algorithm replaces isolated threshold judgment with graph topology, expanding single-point risk into directed risk chains, and the attenuation coefficient is jointly determined by topological distance and spatial sensitivity, so that the propagation results not only conform to the physical distance attenuation law, but also capture near-end connected objects with high spatial sensitivity. The result of the synergistic effect of the three technologies is that the stability of the object detection layer provides reliable geometric input for airspace prediction, the physical constraints of airspace prediction provide credible predicted values ​​and uncertainties for risk index calculation, and the risk propagation algorithm propagates uncertainty and spatial sensitivity together on the graph structure. Ultimately, the system can still output physically credible and traceable risk assessment results under complex conditions where data loss and occlusion coexist.

[0041] It should be noted that under the combined degradation conditions where both meteorological and point cloud data have quality defects, the transmission line clearance prediction algorithm faces the technical challenge of superimposed uncertainties. Specifically, when severe convective weather causes intermittent loss of meteorological sensor data, and dense vegetation in the line corridor causes large-scale occlusion of lidar point clouds, both the geometric input of the catenary fitting and the physical input of the meteorological load degrade simultaneously. Under the traditional single processing path, the absence of either input will cause the prediction result to lose its constraint, resulting in an infinite expansion of the clearance prediction interval or even the output of invalid values. Maintenance personnel cannot extract effective risk decision-making information from this. The above technical problems arise because: traditional methods treat geometric and physical inputs as independent preconditions, lacking a mechanism to compensate for the degradation of either by using constraints in the other direction; point cloud occlusion increases the extraction error of the hanging point coordinates or the lowest point coordinates, causing the catenary fitting residual to exceed the confidence threshold; and the lack of meteorological data prevents the historical compensation mechanism from being activated due to the lack of retrieval anchor points. The two degradations are interlocked, making it impossible to restore prediction capability by repairing only one of the inputs. The usual solutions to the aforementioned technical problems are to introduce redundant sensor arrays or increase sampling density to reduce the probability of obstruction, or to use statistical interpolation methods to fill in missing meteorological data. However, redundant sensor solutions are extremely costly to implement on existing lines, and multiple sensors may be affected simultaneously during severe convective weather, failing to fundamentally eliminate compound degradation conditions. Statistical interpolation methods lack physical constraints, and under extreme conditions, extrapolation errors often exceed safety margins, making it difficult to quantify the uncertainty of the interpolation results and providing a confidence reference for risk decision-making. This invention effectively solves this technical problem, specifically through a collaborative compensation mechanism at the following three levels. In the geometric input layer, the iterative fusion unit uses multi-view geometric consistency constraints to ensure that even if the extraction of the coordinates of the hanging point or lowest point fails due to occlusion in one view, the 3D position of the point can still be recovered through backprojection of the projected coordinates of the object in the other two views. This allows the degradation of the geometric input to be offset by view redundancy rather than being directly passed to the catenary fitting. In the physical input layer, similar working condition retrieval uses temperature, load, and wind speed ranges as anchor points. These three anchor points can come from different sensors. Even if some meteorological sensors are missing, as long as either temperature or wind speed is available, the retrieval can still be performed within the reduced feature space. Although the compensation sample set is reduced, it can still provide physically meaningful estimates of the correction amount. The uncertainty of the compensation result is explicitly output through the width of the prediction interval. At the risk assessment layer, the object relationship graph shortest risk path propagation algorithm accepts the net clearance prediction interval with uncertainty as input. It calculates the risk propagation path separately under the pessimistic scenario corresponding to the lower bound of the interval, so that the risk assessment results under the compound degradation condition are output in the form of conservative estimates rather than invalid values. Operation and maintenance personnel can formulate disposal priorities based on the risk chain under the pessimistic scenario, so as to obtain risk assessment conclusions with decision-making value under the compound degradation condition.

[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 predicting and analyzing the operating conditions of a power transmission line and assessing its risk level.

[0043] A third aspect of the present invention provides a transmission line operating condition prediction analysis and risk level assessment system, comprising the aforementioned computer-readable storage medium. The system is 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 logical self-consistency of the following three levels. First, at the object detection level, after the point cloud data is projected into a multi-view pseudo-image, the geometric projection relationship of the same physical object in different views is determined. Therefore, when the detection confidence of a certain view is low due to occlusion, the projection area of ​​the object in the other two views is still visible. An attention mask can be generated by back-projection of the projection geometry, which strengthens the response of the region in the feature map corresponding to the low-confidence view. Thus, the detection gap can be corrected using existing data without adding additional sensors. This is the geometric basis for the iterative fusion unit to improve the localization stability under occlusion conditions. Secondly, at the level of clearance prediction, the shape of the conductor within the span is constrained by the physical laws of its own weight, temperature, and wind load. The baseline curve given by the catenary model has a clear physical meaning. The thermal elongation correction, equivalent vertical load correction, and lateral wind deflection correction all correspond to the real physical mechanisms. Therefore, the sources of deviation in the prediction curve can be traced item by item. When meteorological or load data is missing, the physical processes experienced by samples with similar temperature, load, and wind speed ranges in the historical operating condition sample library are highly similar to the current operating conditions. Replacing the missing input with the historical correction amount weighted by similarity is equivalent to approximating the current process with a verified physical process. Its error is constrained by similarity rather than random diffusion. Therefore, although the prediction interval width increases when data is missing, the deviation of the prediction center value is physically limited by the historical distribution, and the result is still interpretable. Third, at the risk assessment level, the spatial adjacency and topological connections between objects along the transmission line are objectively real. A conductor segment with insufficient clearance margin will inevitably have a ripple effect on the insulators it is connected to, nearby tree clusters, and the ground sections below. This ripple effect can be quantified by the edge weights of the graph structure. The shortest risk path propagation algorithm searches for risk chains along the direction of decreasing edge weights, and its attenuation coefficient is determined by both topological distance and spatial sensitivity. This ensures that the risk propagation results conform to the physical law that the impact on distant objects decreases, while also identifying nearby objects with high spatial sensitivity. These three levels of mechanisms fill the information gaps introduced by data loss and occlusion from the three dimensions of geometric constraints, physical modeling, and graph topology. They are logically independent yet sequentially connected, enabling the overall scheme to output stable and physically based prediction and assessment results under various incomplete data conditions.

[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 method of step S1 is as follows: Collect three-dimensional laser point cloud data, meteorological data (wind speed, wind direction, temperature), load current data and historical inspection records of the transmission line corridor. Time alignment is performed based on the global navigation satellite system timestamp. The coordinates of each data source are unified to the same geodetic coordinate system. Data quality labels are added to each record. The labels include the completeness of collection, time deviation and coordinate residual.

[0047] The specific implementation of step S2 is as follows: Statistical outlier filtering for noise reduction, voxel downsampling, and octree block indexing are sequentially performed on the original point cloud. Higher point density is retained for regions identified as conductors and insulators, while other regions are appropriately downsampled to construct a multi-scale spatial representation.

[0048] The specific implementation of step S3 is as follows: Project the 3D point cloud along the horizontal, vertical, and forward directions to generate three sets of pseudo-images. Each set of pseudo-images encodes five channels per pixel, representing the height value. (Unit: m) Depth value (Unit: m) Reflection intensity value (Dimensionless) Point Density Value (unit ) and local height difference (Unit: m), forming the pixel feature vector Each component represents the point cloud statistics within that pixel grid. The graph-constrained multi-view pseudo-image detection network takes the three sets of pseudo-images mentioned above as input. The lightweight convolutional backbone of the network front end contains four downsampling stages, each consisting of depthwise separable convolutions and batch normalization. The central direction perception attention module... View number Channel feature map Assign channel weights Spatial weights The enhanced feature map is as follows: ; In the formula, This represents three view indexes. For channel indexing, For the first View number Original feature map of the channel To enhance the feature map, and All are generated by global average pooling of channels, followed by a fully connected layer and The activation function outputs all three values, which have the same dimensions and are all feature activation values. The decoupled detection head outputs the first, second, and third values ​​respectively. Coordinates of the 2D candidate box in the view Category confidence and coordinates of key points within the view ,in The candidate box index is used. The graph consistency constraint layer calculates cross-view association scores for the three view candidate boxes based on projection geometry constraints, category consistency, and topological adjacency. View candidate box With the View candidate box The associated score is: ; In the formula, The score for projective geometric consistency (dimensionless, ranging from 0 to 1). The score is the category consistency score (dimensionless, ranging from 0 to 1). The topological adjacency score (dimensionless, ranging from 0 to 1). , , For the corresponding weight coefficients and satisfying The empirical values ​​were set to 0.5, 0.3, and 0.2 respectively. For dimensionless correlation scoring. The iterative fusion unit iterates 2 to 3 times, the first... In the cycle, if the first View candidate box confidence Then take the remaining views. The candidate regions are used to generate attention masks through coordinate back projection. This is then overlaid onto the feature map corresponding to the low-confidence view, the features are recalculated, and new candidate boxes and confidence scores are output. Regarding resource allocation, the first... Released video memory And assigned to the Memory usage for round attention mask computation satisfy: ; In the formula, For the first The amount of video memory released (in bytes) from the feature map in the middle of the wheel. To be assigned to the The memory usage (in bytes) for round-following attention mask computation, with 0.8 as a fixed allocation ratio (dimensionless), and both sides of the equation having bytes as the unit. View branch feature maps not involved in iterative fusion are marked as read-only and have their memory bandwidth allocation priority reduced during iteration. The network's final output includes object category, 3D candidate space range, and view... Figure 1 Consistency score.

[0049] When establishing the training dataset, 3D point clouds of multiple transmission line corridors were collected and multi-view pseudo-images were generated. Manual annotators labeled each view with 2D bounding boxes and categories. Cross-view association records were created for the bounding boxes of the same physical object across the three views, serving as supervision labels for the graph consistency constraint layer. Low-confidence samples were generated by artificially occluding some views. The training and validation sets were divided in an 8:2 ratio. Network training adopted a phased approach. The first phase only trained the convolutional backbone, attention module, and detection head. The second phase unfroze the graph consistency constraint layer and jointly applied the detection loss and cross-view association loss. The cross-view association loss adopted a contrastive loss form, analyzing the feature vectors of candidate bounding boxes of the same object across different views. and The loss is: ; In the formula, This indicates a pair of objects (bringing the feature distance closer). Indicates different object pairs (pull feature distance). This is the margin hyperparameter (dimensionless), typically ranging from 0.5 to 1.0. For the first View number Feature vectors of candidate boxes The Euclidean distance between the two eigenvectors is dimensionless. This is a dimensionless contrast loss value. Take the larger of the two values ​​within the parentheses. In the third stage, the iterative fusion unit is unfrozen, and the ability to correct low-confidence samples is introduced. The validation set uses detection precision, cross-view association accuracy, and small target recall as comprehensive indicators to determine the final parameters.

[0050] The specific implementation of step S4 is as follows: Point-level segmentation and boundary refinement are performed within the neighborhood of the object-level 3D entity output in step S3. A 3D spatial relationship graph is constructed using objects as nodes and spatial relationships as edges. Node set Includes objects such as towers, conductor segments, insulators, and tree clusters, edge sets Encoding clearance relationships, connection relationships, crossing relationships, proximity relationships, and occlusion relationships.

[0051] The specific implementation of step S5 is as follows. The catenary wind load coupled clearance prediction algorithm uses the coordinates of the hanging point, the coordinates of the lowest point, and the span length extracted from the point cloud as input to fit the catenary model. The static catenary equation is: ; In the formula, For the conductor in a horizontal position Vertical height at the location (in meters). For catenary parameters (in meters). The horizontal tension of the conductor (in N). Mass per unit length of conductor (unit: kg / m). Acceleration due to gravity (unit) (take 9.81). Here are the horizontal coordinates (in meters) of the lowest point of the traverse. The lowest point height (in meters). It is a hyperbolic cosine function. The fitting residuals... The root mean square (in meters) of the difference between the height of each point cloud sampling point and the fitted value represents the basic confidence level. Determined by the following formula: ; In the formula, The value represents the 80th quantile (in meters) of the residual distribution obtained from fitting point clouds of multiple known traverse lines. For dimensionless confidence (values ​​range from 0 to 1), the right side of the equals sign... This is a dimensionless ratio with consistent dimensions. Correction for thermal elongation. (Unit: m) Determined by conductor temperature (Unit: °C) Calculation: ; In the formula, The coefficient of linear expansion of the conductor material (unit: linear expansion coefficient). ), The span length (in meters). The reference temperature is in °C (default 20 °C), and the units on both sides of the equation are in meters (m). Lateral wind displacement. (Unit: m) is: ; In the formula, The lateral wind load per unit length (unit: N / m) is calculated using the following formula: ; In the formula, This is the drag coefficient (dimensionless, empirically taken as 1.0 to 1.2). air density (units) ), Wind speed (unit: m / s) The diameter of the conductor (in meters). Dimensional verification: Dimensions consistent The distance between the gears (in meters). Dimensional verification: The dimensions are consistent. The three types of correction quantities are... , (Sagging increment caused by equivalent vertical load correction, in meters, derived from load current) (Unit: A) Converted to Joule heat increment calculated from material parameters such as conductor resistivity. By superimposing the modified traverse prediction curve onto the catenary baseline curve, a minimum distance search is performed on the point clouds of the entire length with respect to the ground, tree barriers, buildings, and intersections to obtain the predicted clearance value. (Unit: m). When meteorological or load data is missing, similar operating condition samples are retrieved from the historical database. Similarity weights of individual samples satisfy: ; In the formula, For the first The similarity score (dimensionless) between historical samples and the current operating conditions is calculated by weighting the distance between temperature, load, and wind speed intervals. The threshold is determined by the lower limit of the mean similarity within the same cluster in historical sample cluster analysis. For dimensionless weights, both sides of the equation have the same dimensions. Prediction interval width. (Unit: m) is: ; In the formula, This is the proportionality coefficient (unit: m). To compensate for the sample size (dimensionless positive integer). For dimensionless basic confidence, the dimension on both sides of the equation is m. The lower or The less The larger.

[0052] The specific implementation of step S6 is as follows. The object relationship graph shortest risk path propagation algorithm uses a three-dimensional spatial relationship graph. For input, each edge Assigning weights to four categories to form a weight vector ,in Safety margin weight (dimensionless). The distance change rate weight (dimensionless). Spatial sensitivity weight (dimensionless). To measure the mass weight (dimensionless). Edge Multifactor risk index (Dimensionless) is: ; In the formula, , , , The normalized quantified values ​​(dimensionless) of the corresponding factors are all calculated by combining the net airspace prediction value output from step S5 with the confidence interval and the safety threshold. The sum of all weighted terms, with consistent dimensions. Identification. Edges with a risk level higher than the preset level are designated as risk source edges, and the risk propagates from the risk source edge to adjacent nodes. Cumulative risk value during multi-level propagation (Dimensionless) is: ; In the formula, The initial risk index (dimensionless) is the risk index at the edge of the risk source. For the first Level attenuation coefficient (dimensionless, range of values) ), by the topological distance between adjacent edges Spatial sensitivity The formula is determined together as follows: ; In the formula, , The weighted coefficient is a dimensionless coefficient determined by retrospective statistics of the propagation paths of historical risk events. For the first The topological distance of the propagation (dimensionless, in terms of the number of edges). This represents the spatial sensitivity weight (dimensionless), with all terms on the right-hand side being dimensionless and having consistent dimensions. When the cumulative risk value of an adjacent edge exceeds its safety margin threshold, it is included in the risk propagation path and continues to propagate until the cumulative risk value falls below the cutoff threshold or reaches a graph boundary node. All propagation paths are sorted according to their cumulative risk weights to identify risk sources and affected objects. The priority of handling each node on the path is determined by its weight, and the multi-factor risk index and risk level are output.

[0053] To better understand and implement this invention, the following is a specific application scenario of this invention, Example 2: The test scenario used a typical span of a 220kV double-circuit transmission line corridor. This span was 450m long and contained a high-density tree-covered area. During the test, the weather station experienced approximately two hours of missing wind speed data. The test objective was to verify the ability of this invention to predict airspace and assess risk under conditions of tree cover and partial missing meteorological data.

[0054] First, data acquisition and preprocessing were completed. An airborne lidar flew along the conductor corridor to collect 3D point clouds, with an initial density of approximately 80 points per square meter. Meteorological data, load current data, and temperature data were collected simultaneously, and historical operating conditions for the same span over the past three years were extracted from the historical inspection database. After time alignment, data quality was flagged, with wind speed data showing missing measurements during the testing period, which was approximately 12% of the total acquisition time. After statistical filtering and denoising, approximately 3.2% of noise points in the conductor corridor area were removed. Higher point density was retained for the semantic regions of conductors and insulators at smaller voxel sizes, resulting in an average point density of approximately 15 points per meter for the conductor segment after downsampling. After multi-scale processing, each point was associated with a five-dimensional descriptor for pseudo-image generation.

[0055] Subsequently, multi-view pseudo-images are generated and object detection is performed. Three sets of five-channel pseudo-images are generated by projecting along the line direction, perpendicular to the line direction, and vertically, respectively, as follows: Figure 2 As shown, Figure 2 The projection distribution of three pseudo-image views onto towers, conductors, and tree obstructions within the test span is shown. A significant projection overlap between the tree obstruction area and the conductor segment is observed in the top-view view. Inputting the three sets of pseudo-images into a trained map-constrained multi-view pseudo-image detection network, in the tree obstruction area, the confidence score of the corresponding conductor candidate box in the top-view view is 0.42, lower than the low confidence threshold of 0.5, triggering an iterative fusion unit. The confidence scores of the corresponding candidate boxes in the longitudinal and transverse line views are 0.81 and 0.78, respectively, both higher than the high confidence threshold of 0.7. After coordinate back-projection to generate an attention mask, it is superimposed on the top-view feature map. After two iterations, the confidence score of this candidate box increases to 0.73, and the 3D entity localization of the object is completed. Finally, 2 towers, 4 conductor segments, 6 insulator strings, 1 tree obstruction cluster, and 1 crossing point are detected, providing a complete 3D entity representation at the object level.

[0056] After object detection, point-level segmentation and boundary refinement are performed in the neighborhood of each object. A region growing segmentation method is used, with the normal vector similarity threshold set to 15° and the curvature continuity threshold set to 0.05. A three-dimensional spatial relationship graph is constructed, generating 14 graph nodes and 23 graph edges. The edge type distribution is shown in Table 1.

[0057] Table 1. Distribution of Edge Types in 3D Spatial Relationship Graph

[0058] The catenary-wind load coupled clearance prediction algorithm was used to calculate the clearance for each of the four conductor segments sequentially. Taking the third conductor segment, which suffered the most severe shading, as an example, the catenary fitting residual was 0.083m, lower than the basic confidence threshold (0.12m based on the 80th percentile of historical fitting experiments), resulting in a high basic confidence rating. With a temperature of 32℃, a load current of 680A, and missing wind speed data, the algorithm retrieved 14 similar operating condition samples from the historical database, covering a temperature range of 28-36℃ and a load range of 600-750A. The similarity-weighted average lateral wind deflection correction was set at 0.38m. After superimposing the thermal elongation correction, the equivalent vertical load correction, and the lateral wind deflection correction, the minimum ground clearance prediction value for the third conductor segment was 5.21m, with a prediction range of 4.87-5.55m. The predicted lateral clearance of the third conductor segment to the tree-blocking cluster is 2.14m, with a prediction range of 1.93 to 2.35m, which is below the safety threshold of 3m. The predicted clearance results for each conductor segment are summarized in Table 2.

[0059] Table 2 Summary of the predicted clearance results for each section of the conductor

[0060] Using the object relationship graph shortest risk path propagation algorithm, the edge with the highest risk index between the third guide line and the tree barrier cluster was identified as the risk source edge. Figure 3 As shown, Figure 3 The visualization of the risk propagation path on a 3D spatial relationship map is presented. The risk originates from the clearance relationship edge between the third conductor segment and the tree cluster, propagates along the connection edge to the two insulator nodes connected to the third conductor segment, and then propagates along the adjacent relationship edge to the ground partition node. During propagation, the risk value decreases according to the attenuation coefficient. The two insulator strings corresponding to the third conductor segment are included in the risk propagation path. When the cumulative risk value of the ground partition node falls below the cutoff threshold, the propagation terminates. The ranking results determine that the risk source is the third conductor segment and the tree cluster, and the affected objects are the two insulator strings connected to the third conductor segment. The highest priority for handling is the tree clearing operation in the corridor of the third conductor segment, followed by the routine inspection of the two insulator strings. The risk evolution result is output through object identity inheritance, recording the risk level of this span in this assessment as level three (higher risk), and comparing it with the historical evolution sequence, confirming that the risk level of this span has been increasing in the past three months. The system outputs corresponding handling suggestions.

[0061] Compared with traditional methods, the technological advancements brought by this invention are reflected in the following aspects: Traditional single-view point cloud detection cannot complete the 3D positioning of the conductor in areas obstructed by tree barriers, requiring manual supplementary measurements. However, this invention, through multi-view geometric consistency iterative correction, transforms the obstruction problem into a problem of information complementarity between views, enabling positioning without additional data collection. Traditional catenary calculation can only stop calculation or use fixed values ​​when meteorological data is missing. However, this invention, through a physical similarity weighted compensation mechanism, ensures that the impact of missing data on the predicted central value is constrained by historical physical processes, and the prediction uncertainty is explicitly output in interval form. Traditional risk assessment mainly relies on single-point threshold judgment, which cannot trace the source and spread of risk. However, this invention, through graph topology propagation, extends single-point risks into directed risk chains, allowing maintenance personnel to directly obtain the basis for handling priorities from the propagation path, realizing a leap from result output to decision support.

[0062] It should be noted that the variables involved in this invention are explained in detail in Tables 3 and 4.

[0063] Table 3. Variable Explanation Table (Part 1)

[0064] Table 4. Variable Explanation Table (Part Two)

[0065] 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 predicting and analyzing the operating conditions and assessing the risk level of transmission lines, characterized in that, Includes the following steps: Collect 3D point cloud data, meteorological data, load data, temperature data and historical inspection data of the transmission line corridor, perform time alignment and coordinate benchmark unification, and generate data quality labels; The 3D point cloud data is processed by denoising, downsampling, block indexing and multi-scale spatial representation, and a high point density is retained for conductor and insulator regions. Generate multi-view pseudo-images, input a graph-constrained multi-view pseudo-image detection network, and output object-level 3D entity representations of towers, conductors, insulators, tree barriers, and crossings; Point-level segmentation and boundary refinement are performed within the neighborhood of object-level 3D entities to construct a 3D spatial relationship graph with objects as nodes and spatial relationships as edges; Based on the three-dimensional spatial relationship map, the catenary-wind load coupled clearance prediction algorithm is called to calculate the characteristics of conductor clearance, sag, wind deflection and icing load conditions, as well as future predicted values ​​and confidence intervals. Based on the predicted working condition characteristics, measurement quality score and safety threshold, the object relationship graph shortest risk path propagation algorithm is invoked to calculate and output the multi-factor risk index and risk level. Output risk evolution results and disposal suggestions through object identity inheritance.

2. The method for predicting and analyzing the operating conditions and assessing the risk level of transmission lines according to claim 1, characterized in that, The input of the graph-constrained multi-view pseudo-image detection network specifically receives three sets of pseudo-images: a forward line view, a horizontal line view, and a top view. Each pixel channel encodes five channels: height value, depth value, reflection intensity value, point density value, and local height difference value.

3. The method for predicting and analyzing the operating conditions and assessing the risk level of transmission lines according to claim 2, characterized in that, The network front-end of the graph-constrained multi-view pseudo-image detection network specifically consists of a lightweight convolutional backbone with four downsampling stages. Each stage is composed of depthwise separable convolutional layers and batch normalization layers. An orientation-aware attention module is embedded in the middle to assign different channel weights and spatial weights to conductor features, insulator string features, and tower features, respectively.

4. The method for predicting and analyzing the operating conditions and assessing the risk level of transmission lines according to claim 3, characterized in that, The detection head of the graph-constrained multi-view pseudo-image detection network adopts a decoupled structure, outputting two-dimensional candidate box coordinates, category confidence, and key point coordinates within the view respectively. After the detection head, a graph consistency constraint layer is set to perform soft matching on the candidate boxes output by the three views according to projection geometric constraints, category consistency, and topological adjacency, generating cross-view association scores.

5. The method for predicting and analyzing the operating conditions and assessing the risk level of transmission lines according to claim 4, characterized in that, The graph-constrained multi-view pseudo-image detection network is equipped with an iterative fusion unit. The iterative fusion unit iterates for a preset number of rounds. In each round, if the confidence of a candidate box in a certain view is lower than the low confidence threshold, then the candidate regions in the other two views with confidence higher than the high confidence threshold are selected. An attention mask is generated by coordinate back projection and superimposed on the feature map corresponding to the low confidence view. The features of the view are recalculated and the candidate box and confidence are re-output.

6. The method for predicting and analyzing the operating conditions and assessing the risk level of transmission lines according to claim 5, characterized in that, When meteorological or load data is missing, the catenary-wind load coupled headroom prediction algorithm retrieves similar operating condition samples with similar temperature, load, and wind speed ranges from the historical database. The weighted average of the correction values ​​is used as compensation, with weights allocated from high to low similarity. The similarity threshold is determined by cluster analysis of historical samples. The prediction interval width is determined by a combination of basic reliability and the number of compensation samples.

7. The method for predicting and analyzing the operating conditions and assessing the risk level of transmission lines according to claim 6, characterized in that, The training of the graph-constrained multi-view pseudo-image detection network adopts a phased training method: in the first phase, the parameters of the graph consistency constraint layer and the iterative fusion unit are fixed, and only the convolutional backbone, attention module and detection head are trained; The second stage unfreezes the map consistency constraint layer and uses cross-view association loss in the form of contrastive loss for joint training. The third stage unfreezes the iterative fusion unit, introducing the ability of low-confidence samples to train the network and utilize high-confidence views to correct the features of low-confidence views.

8. The method for predicting and analyzing the operating conditions and assessing the risk level of transmission lines according to claim 7, characterized in that, The aforementioned catenary-wind load coupled clearance prediction algorithm specifically uses the coordinates of the hanging point, the coordinates of the lowest point, and the span length extracted from the point cloud to fit the catenary model. The fitting residual is used to generate the basic reliability. The thermal elongation correction is calculated based on the conductor temperature data, the equivalent vertical load correction is calculated based on the line load current and conductor material parameters, and the lateral wind displacement is calculated based on the wind speed and direction data. The three types of corrections are superimposed on the catenary baseline curve to obtain the corrected conductor prediction curve.

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 predicting and analyzing the operating conditions of a transmission line and assessing its risk level as described in any one of claims 1-8.

10. A system for predicting and analyzing the operating conditions and assessing the risk level of 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