AI-based tree barrier threat level assessment method and device

Abstract

The embodiment of the application provides a kind of based on AI's tree barrier threat level evaluation method and device, through space analysis and three-dimensional reconstruction, realize the accurate modeling of tree barrier.Construct evaluation mechanism, combined with comprehensive evaluation and dynamic evaluation, establish reliable threat early warning strategy.Introduce control optimization, through early warning prediction and hierarchical control, ensure the continuous safety of line.The method effectively solves the deficiency of traditional technology in tree barrier modeling, threat evaluation and early warning control, etc., provides technical support for transmission line safety management.

Description

Technical Field

[0001] This application relates to the field of data processing, specifically to an AI-based method and apparatus for assessing tree barrier threat levels. Background Technology

[0002] Existing tree barrier threat assessment methods have significant shortcomings. Traditional systems perform poorly in morphological parameter extraction and spatial analysis, failing to effectively and accurately model tree barriers and thus affecting assessment results.

[0003] Furthermore, existing technologies suffer from bottlenecks in hazard calculation and risk scoring. Most systems lack a comprehensive evaluation mechanism and dynamic assessment strategy, resulting in insufficient accuracy in threat level determination.

[0004] Existing systems have technical shortcomings in early warning and control. They lack in-depth analysis of growth trends, making it difficult to achieve efficient early warning and prediction through collaborative training, thus affecting control effectiveness. Solving these problems is of great significance for improving the safety management capabilities of transmission lines. Summary of the Invention

[0005] To address the problems in existing technologies, this application provides an AI-based method and device for assessing tree obstacle threat levels, which can effectively solve the shortcomings of traditional technologies in tree obstacle modeling, threat assessment, and early warning and control, and provide technical support for the safety management of power transmission lines.

[0006] To solve at least one of the above problems, this application provides the following technical solution:

[0007] Firstly, this application provides an AI-based method for assessing tree barrier threat levels, including:

[0008] Based on the tree canopy morphology parameter set and line safety parameter set acquired from high-definition images of transmission lines, a coordinate mapping table is constructed according to the safety distance regulations for power facilities. The coordinate mapping table is used to generate a tree barrier spatial distribution model through topological analysis. The tree barrier spatial distribution model is then reconstructed in three dimensions to obtain the protection area division result. Based on the protection area division result, tree canopy projection range data is extracted. The tree canopy projection range data is then correlated with the tree canopy morphology parameter set to generate a growth trend set.

[0009] A spatial approximate vector set is constructed from the tree canopy projection range data. Distance calculation is performed on the spatial approximate vector set to obtain a group of risk level indicators. Fuzzy comprehensive evaluation is performed on the group of risk level indicators to generate a risk score matrix. A tree barrier level label set is obtained by hierarchical mapping based on the risk score matrix. The tree barrier level label set and the growth trend set are co-trained to obtain a dynamic evaluation model.

[0010] The dynamic assessment model is deployed in a real-time monitoring system to receive tree obstacle monitoring data and make predictions to obtain threat warning results. The threat warning results are correlated with historical inspection records to generate warning instructions. Based on the warning instructions, the tree obstacles on the transmission line are managed in a hierarchical manner.

[0011] Furthermore, it also includes: extracting the vertical projection area and spatial height data of the tree canopy based on high-definition images of the transmission line; normalizing the vertical projection area and spatial height data of the tree canopy according to the safety distance regulations for power facilities to obtain a tree canopy morphology parameter set; generating a line safety parameter set according to the network topology structure from the substation to the transmission line; and performing coordinate system transformation between the tree canopy morphology parameter set and the line safety parameter set to obtain a spatial feature matrix.

[0012] The spatial feature matrix is ​​divided into spatial grids to generate a three-dimensional coordinate mapping table. The distance vector between the tree obstacle location information in the three-dimensional coordinate mapping table and the boundary of the route channel is calculated to obtain a coordinate mapping table. Based on the coordinate mapping table, a topological connectivity relationship is constructed to obtain a tree obstacle spatial distribution model.

[0013] Furthermore, it also includes: subdividing the tree barrier spatial distribution model into a grid according to the boundary constraints of the power facility protection area to obtain a spatial point cloud; performing three-dimensional surface fitting on the spatial point cloud to generate a voxel model; constructing a spatial occupancy probability map based on the voxel model to obtain the region division result; and calculating the minimum bounding rectangle of the region division result to obtain the tree canopy projection range data.

[0014] The growth feature sequence is obtained by performing time-series registration on the tree canopy projection range data. The growth feature sequence is then mapped in multiple dimensions to the tree canopy morphology parameter set to obtain a dynamic change matrix. A growth rate model is then established based on the dynamic change matrix to generate a growth trend set.

[0015] Furthermore, it also includes: performing feature dimensionality reduction processing on the tree canopy projection range data to obtain a principal component feature set; performing projection transformation on the principal component feature set and the centerline of the line channel to obtain a distance vector set; constructing a K-nearest neighbor search tree based on the distance vector set to obtain a spatial approximate vector set; and calculating the danger level index set by Euclidean distance according to the boundary of the power facility protection area.

[0016] The risk level index group is mapped to a fuzzy evaluation set according to the safety level threshold. The fuzzy evaluation set is weighted based on a preset weight coefficient to obtain a comprehensive score vector. The comprehensive score vector is normalized with historical evaluation data to generate a risk score matrix.

[0017] Furthermore, it also includes: dividing the risk scoring matrix hierarchically according to hierarchical clustering rules to obtain risk level clusters; generating an initial label set by label encoding the risk level clusters based on safety distance constraints; verifying the association between the initial label set and historical control records to obtain a tree barrier level label set; and constructing a training sample library based on the tree barrier level label set.

[0018] The label data in the training sample library and the growth trend set are fused to obtain a training feature matrix. The training feature matrix is ​​then input into a deep neural network for supervised learning to obtain a prediction model. The prediction model is then cross-validated and its parameters are optimized to generate a dynamic evaluation model.

[0019] Furthermore, it also includes: converting the dynamic evaluation model into a lightweight inference engine to obtain a prediction computing unit, performing model quantization and pruning optimization on the prediction computing unit to obtain a real-time inference model, deploying the real-time inference model to the edge computing node of the monitoring system, and constructing a data preprocessing pipeline based on the edge computing node to obtain a real-time analysis unit.

[0020] The real-time analysis unit receives tree obstacle monitoring data and extracts features to obtain a monitoring feature vector. The monitoring feature vector is then input into the real-time inference model to predict the threat level and obtain a predicted probability distribution. The predicted probability distribution is then used to determine a threshold according to the early warning rules to generate a threat early warning result.

[0021] Furthermore, it also includes: aligning threat warning results with historical inspection records according to spatiotemporal indexes to obtain a related dataset; performing time-series correlation analysis on the related dataset to obtain risk evolution characteristics; constructing a warning rule base based on the risk evolution characteristics to obtain a warning strategy model; and performing rule matching between the warning strategy model and real-time warning data to generate warning instructions.

[0022] Based on the threat level of the warning instruction, a control priority queue is established to obtain a task scheduling table. The task scheduling table is partitioned according to the operation resource constraints to obtain an operation assignment scheme. Based on the operation assignment scheme, differentiated control measures are implemented for tree obstacles on transmission lines and the control results are recorded.

[0023] Secondly, this application provides an AI-based tree barrier threat level assessment device, comprising:

[0024] The tree obstacle identification module is used to acquire tree crown morphology parameter sets and line safety parameter sets based on high-definition images of transmission lines, construct a coordinate mapping table according to the power facility safety distance regulations, generate a tree obstacle spatial distribution model through topological analysis of the coordinate mapping table, perform three-dimensional reconstruction of the tree obstacle spatial distribution model to obtain the protection area division results, extract tree crown projection range data based on the protection area division results, and perform correlation analysis between the tree crown projection range data and the tree crown morphology parameter set to generate a growth trend set;

[0025] The tree barrier analysis module is used to construct a spatial approximate vector set from the tree canopy projection range data, perform distance calculation on the spatial approximate vector set to obtain a risk level index group, perform fuzzy comprehensive evaluation on the risk level index group to generate a risk score matrix, perform hierarchical mapping based on the risk score matrix to obtain a tree barrier level label set, and perform collaborative training of the tree barrier level label set and the growth trend set to obtain a dynamic evaluation model.

[0026] The risk warning module is used to deploy the dynamic assessment model on the real-time monitoring system, receive tree obstacle monitoring data and make predictions to obtain threat warning results, correlate the threat warning results with historical inspection records to generate warning instructions, and perform hierarchical management and control of tree obstacles on transmission lines according to the warning instructions.

[0027] Thirdly, this application provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the AI-based tree barrier threat level assessment method.

[0028] Fourthly, this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the AI-based tree barrier threat level assessment method.

[0029] Fifthly, this application provides a computer program product, including a computer program / instructions that, when executed by a processor, implement the steps of the AI-based tree obstacle threat level assessment method.

[0030] As described above, this application provides an AI-based method and apparatus for assessing tree obstacle threat levels. Through spatial analysis and 3D reconstruction, it achieves accurate modeling of tree obstacles. An assessment mechanism is constructed, combining comprehensive evaluation and dynamic assessment to establish a reliable threat early warning strategy. Control optimization is introduced, ensuring the continuous safety of power lines through early warning prediction and tiered control. This method effectively addresses the shortcomings of traditional technologies in tree obstacle modeling, threat assessment, and early warning control, providing technical support for the safety management of transmission lines. Attached Figure Description

[0031] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0032] Figure 1 This is a flowchart illustrating the AI-based tree obstacle threat level assessment method in the embodiments of this application;

[0033] Figure 2 This is a structural diagram of the AI-based tree obstacle threat level assessment device in the embodiments of this application; Detailed Implementation

[0034] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0035] The acquisition, storage, use, and processing of data in this application comply with relevant laws and regulations.

[0036] In view of the problems existing in the prior art, this application provides an AI-based method and device for assessing the threat level of tree obstacles. Through spatial analysis and 3D reconstruction, it achieves accurate modeling of tree obstacles. An assessment mechanism is constructed, combining comprehensive evaluation and dynamic assessment to establish a reliable threat early warning strategy. Control optimization is introduced, ensuring the continuous safety of transmission lines through early warning prediction and hierarchical control. This method effectively solves the shortcomings of traditional technologies in tree obstacle modeling, threat assessment, and early warning control, providing technical support for the safety management of transmission lines.

[0037] To effectively address the shortcomings of traditional technologies in tree obstacle modeling, threat assessment, and early warning and control, and to provide technical support for the safety management of transmission lines, this application provides an embodiment of an AI-based tree obstacle threat level assessment method. See [link to embodiment]. Figure 1 The AI-based tree barrier threat level assessment method specifically includes the following:

[0038] Step S101: Based on the high-definition images of the transmission line, acquire the tree canopy morphology parameter set and the line safety parameter set, construct a coordinate mapping table according to the power facility safety distance regulations, generate a tree barrier spatial distribution model through topological analysis of the coordinate mapping table, perform three-dimensional reconstruction of the tree barrier spatial distribution model to obtain the protection area division result, extract the tree canopy projection range data based on the protection area division result, and perform correlation analysis between the tree canopy projection range data and the tree canopy morphology parameter set to generate a growth trend set;

[0039] First, based on high-definition images of transmission lines, tower locations, conductor routes, and vegetation area masks are read to extract initial elements of the canopy morphology parameter set. Specifically, semantic segmentation is performed on single-frame images to obtain the canopy boundaries, and then stereo calibration is used to recover the canopy spatial height and vertical projected area, while simultaneously extracting the canopy edge curvature and center of gravity position. Simultaneously, the corridor centerline, boundary zone, and minimum safety distance clauses are read from line design data and survey data, organized into a line safety parameter set, and aligned within the same survey area coordinate system as input for subsequent mapping.

[0040] Based on the aforementioned inputs, the image coordinates and ground coordinates are registered according to the reference points in the procedure, generating a coordinate mapping table. During processing, the tower foundation points and conductor suspension points are used as the control point set, and the tree canopy boundary pixels are back-projected in batches into three-dimensional discrete points. For points with abrupt changes across frames, time interpolation is performed and the data source is labeled to ensure the stability and traceability of subsequent connectivity determinations.

[0041] Based on the coordinate mapping table, a topological analysis is performed to obtain a spatial distribution model of tree barriers. The method involves establishing a ternary index of longitudinal mileage, lateral distance, and height along the channel centerline, connecting neighboring points according to the nearest neighbor rule to form a canopy segment subgraph; then, combining the line segmentation information to supplement cross-frame edge connections, obtaining the spatial connectivity structure covering individual trees and communities. The nearest distance and direction sign to the channel boundary are registered for each connected component, providing constraints for subsequent region division and trend analysis.

[0042] Based on the aforementioned tree barrier spatial distribution model, a 3D reconstruction is performed, and the protected area delineation results are output. Connected components are converted into point clouds, and voxel occupancy representations are obtained using local surface fitting. Subsequently, spatial intersection calculations are performed with the protocol boundaries, and voxels are labeled with three categories: inner, boundary zone, and outer, along with their corresponding connected components. These delineation labels are then backfilled into the component attributes, serving as boundary conditions for projection and temporal registration.

[0043] Based on the protected area delineation results, tree canopy projection range data is extracted. Connected components are projected onto the ground plane, and the minimum bounding rectangle and principal axis direction are calculated to obtain the projected area, the angles between the major and minor axes and the principal axis relative to the channel centerline, and the association with the area label is retained. For multi-period images, temporal matching of the same object is established through centroid and principal axis similarity, forming a temporally aligned projection sequence.

[0044] Based on the tree canopy projection range data, and combined with the tree canopy morphology parameter set, trend determination is completed and a growth trend set is generated. To ensure that the rules can be uniformly applied, this embodiment provides a scoring formula:

[0045] Q = p·u + q·v − r·w

[0046] In the formula, Q is the overall trend score; p, q, and r are non-negative weights; u is the time variation term of the projected area, taking the incremental ratio between adjacent observations; v is the component intensity of the lateral vector in the inner direction of the channel; w is the relative deviation term between height and lateral, representing the substitution effect of height growth on lateral approximation. When Q is positive and v is positive, it is labeled as a channel approximation trend; when Q is negative and w is dominant, it is labeled as a vertically dominant trend; when both are close to the tradeoff boundary, it is registered as a mixed trend, and the time window and the index of participating observations are recorded. The weights are set by the robust statistical interval of historical samples and do not involve fixed values.

[0047] Based on the growth trend set, trend labels and corresponding windows are written back to the node and edge attributes of the tree barrier spatial distribution model to form traceable time annotations. Simultaneously, the projection range data and region labels are organized into distance and direction input blocks for constructing downstream hazard level indicators. Trend labels, as dynamic priors, are read in subsequent hierarchical mapping and collaborative training stages.

[0048] Finally, the growth trend set and projection range data are linked to a coordinate mapping table via a reference key, so that they can be directly called by the dynamic evaluation model in subsequent steps. Specifically, the trend comprehensive score Q will be used as an auxiliary feature of the training samples in the generation of the risk scoring matrix, and will be used in the real-time monitoring stage to limit the direction and rate of short-term predictions, thereby maintaining the temporal consistency and data replayability of the evaluation chain.

[0049] Step S102: Construct a spatial approximate vector set from the tree canopy projection range data, perform distance calculation on the spatial approximate vector set to obtain a risk level index group, perform fuzzy comprehensive evaluation on the risk level index group to generate a risk score matrix, perform hierarchical mapping based on the risk score matrix to obtain a tree barrier level label set, and perform collaborative training between the tree barrier level label set and the growth trend set to obtain a dynamic evaluation model.

[0050] First, the aforementioned canopy projection range data and protected area labels are read. Three descriptions—longitudinal mileage, lateral distance, and principal axis angle—are established according to the channel centerline and aligned with the spatial height of the canopy morphology parameter set to form an input block for neighborhood retrieval. To reduce redundancy, segments with similar heights in adjacent time slices of the same connected component are merged with stable segments of the principal axis. Keyframe indices representing morphology and orientation are retained as reference samples for subsequent construction.

[0051] Based on the input block, a spatial approximation vector set is constructed. Specifically, each keyframe sample is encoded as a four-element vector, containing lateral distance, longitudinal odometer difference, projected major axis length, and principal axis orientation sign, and the corresponding protected area label is registered. A hierarchical index is established for all samples, first clustered by area label, and then a two-dimensional grid of odometer and lateral distance is used within each cluster for fast lookup to ensure that distance calculations are performed only within physically relevant ranges.

[0052] Based on the aforementioned spatial approximation vector set, distance calculations are performed to obtain a set of hazard level indicators. The calculations include the lateral distance to the channel boundary, the connectivity distance between adjacent samples along the line, and the vertical margin to the lowest point of the guide wire. A stability penalty term is added to samples falling within the boundary zone. To unify subsequent evaluations, the three types of distances and penalties are normalized to the same scale, and an indicator entry containing distance, orientation, and stability labels is generated for each sample, serving as direct input for fuzzy evaluation.

[0053] Based on the aforementioned risk level index set, a fuzzy comprehensive evaluation is performed to generate a risk scoring matrix. The method involves setting three membership functions according to the safety distance protocol, mapping the risk membership degrees of lateral, longitudinal connectivity, and vertical margin respectively, and then weighting the results using source reliability as the weight to output a scoring table for the samples and risk dimensions. The scoring table also retains the index of the index source and keyframe, facilitating closed-loop traceability with previous mapped data.

[0054] Based on the risk scoring matrix, a hierarchical mapping is performed to obtain a tree barrier level label set. During processing, the scoring table is first divided into several clusters using hierarchical clustering. Then, each cluster is assigned a safety level code based on the protected area label, and the code is solidified into a label after consistency verification with historical control records. The label entries use keyframe indexes as the primary key, along with inter-cluster boundary and uncertainty area markers, to identify transitional samples during the training phase.

[0055] Based on the aforementioned tree barrier level label set, a dynamic evaluation model is obtained through co-training with the aforementioned growth trend set. During training, the risk dimension of the scoring matrix and the direction and rate of trend entries are used as joint features. A supervised learning structure called the Tree Barrier Dynamic Discriminant Network is employed, with keyframe sample sequences and corresponding labels as inputs and level probability distributions as outputs. During training, samples in the uncertainty zone are assigned lower weights, and the growth trend direction is used as a consistency constraint for short-term predictions to ensure that the model has temporal continuity in distinguishing between approximation-type and vertically dominant samples.

[0056] Finally, the dynamic assessment model, along with the risk scoring matrix and label mapping relationship, is registered as a callable entity for direct loading in subsequent real-time monitoring steps. Specifically, the model can read the keyframe index to trace back to the projection range data and spatial approximation vector, which is used for input organization in online prediction and threshold determination, ensuring field consistency and path replayability from offline training to online inference.

[0057] Step S103: Deploy the dynamic evaluation model in the real-time monitoring system, receive tree obstacle monitoring data and make predictions to obtain threat warning results, correlate the threat warning results with historical inspection records to generate warning instructions, and perform hierarchical management and control of tree obstacles on transmission lines according to the warning instructions.

[0058] First, the label mapping relationship between the aforementioned dynamic assessment model and the risk scoring matrix is ​​read, and the model is converted into a lightweight inference unit adapted to edge nodes. This unit is then aligned with fields of the channel centerline, protected area labels, and keyframe indexes to form a real-time input interface. After accessing tree obstacle monitoring data, time alignment and coordinate registration are performed on the image stream and laser point cloud, respectively. Monitoring feature vectors are generated according to a predetermined window, with fields including lateral distance, projection principal axis, spatial height, and stability markers, while retaining the source device and timestamp.

[0059] Based on the real-time input interface, monitored feature vectors are batch-fed into the dynamic evaluation model, outputting a level probability distribution and trend consistency score. During the inference phase, keyframes are used to backtrack to the dimension weights in the risk scoring matrix, reducing the confidence level for short-term mutation samples and completing threshold determination within edge nodes, generating a threat warning result containing the level, triggering dimension, and evidence pointers. This result also registers a protected area label for subsequent compliance checks.

[0060] Based on the threat warning results, historical inspection records are loaded and aligned by channel mileage and time to form associated data blocks. Temporal correlation analysis is performed on these associated data blocks to extract persistence, recurring locations, and seasonal characteristics. These characteristics are then mapped to the warning rule base to generate executable warning commands. The warning commands clearly define the target mileage segment, object components, and suggested time limits, and include triggering evidence to ensure traceability.

[0061] Based on the aforementioned early warning instructions, a tiered management queue is constructed. During processing, tasks are categorized into three types—urgent cleanup, time-limited re-inspection, and routine observation—according to the level of the warning and the protected area label. A preliminary schedule is then provided in conjunction with a resource occupancy table. For objects connected across segments, tasks are merged and dispatched in order of upstream proximity to avoid duplicate operations and blind spots in observation.

[0062] Based on the tiered control queue, dispatch entries are generated for the work side. These entries include mileage coordinates, arrival distance, work window and safety boundary prompts, and photo collection requirements for on-site verification. After dispatch, the real-time system continuously receives transmitted images and location information, establishes a closed-loop status record for the same object, and automatically escalates the warning instruction and prompts for a second verification if the risk level does not decrease within two consecutive windows.

[0063] Based on the closed-loop status record, newly generated monitoring features are compared with historical entries. Significant differences trigger secondary inference to avoid single-observation errors. Secondary inference uses the dynamic evaluation model and reads the trend consistency score to suppress short-term changes with conflicting directions. If the result is stable, it is fixed as the latest status, and the execution time limit of the warning instruction is updated synchronously.

[0064] Finally, the execution results of the hierarchical control are written back to the auxiliary labels of the training samples for retraining and threshold calibration in the subsequent offline phase. A one-to-one correspondence is maintained between threat warning results, warning instructions, and execution results, providing direct access to subsequent strategy generation and resource orchestration modules, achieving seamless integration from online prediction and rule-based decision-making to on-site handling.

[0065] As described above, the AI-based tree obstacle threat level assessment method provided in this application can achieve accurate tree obstacle modeling through spatial analysis and 3D reconstruction. An assessment mechanism is constructed, combining comprehensive evaluation and dynamic assessment to establish a reliable threat early warning strategy. Control optimization is introduced, ensuring the continuous safety of transmission lines through early warning prediction and hierarchical control. This method effectively addresses the shortcomings of traditional technologies in tree obstacle modeling, threat assessment, and early warning control, providing technical support for the safety management of transmission lines.

[0066] In one embodiment of the AI-based tree obstacle threat level assessment method of this application, it may further include the following:

[0067] Step S201: Extract the vertical projection area and spatial height data of tree canopy from the high-definition image of the transmission line, normalize the vertical projection area and spatial height data of tree canopy according to the safety distance regulations for power facilities to obtain the tree canopy morphology parameter set, generate the line safety parameter set according to the network topology structure from the substation to the transmission line, and perform coordinate system transformation between the tree canopy morphology parameter set and the line safety parameter set to obtain the spatial feature matrix.

[0068] Step S202: Divide the spatial feature matrix into a spatial grid to generate a three-dimensional coordinate mapping table. Calculate the distance vector between the tree obstacle location information in the three-dimensional coordinate mapping table and the boundary of the route channel to obtain the coordinate mapping table. Based on the coordinate mapping table, construct the topological connectivity relationship to obtain a tree obstacle spatial distribution model.

[0069] First, high-resolution images of the transmission line are accessed, and tree canopy and background segmentation is performed on single frames to extract canopy boundaries and centroids. Combined with stereo image pairs or multi-view constraint adjustments, spatial height is calculated, and the vertical projected area on the ground surface is obtained, forming two original sequences of height and area data. To compensate for differences in imaging scale and viewpoint distortion, based on the reference scale and minimum safety clearance in the power facility safety distance regulations, interval normalization and outlier segment labeling are performed on the two sequences to obtain a set of tree canopy morphology parameters, including normalized height, normalized projected area, and boundary confidence markers.

[0070] Based on the aforementioned canopy morphology parameter set, the network topology from the substation to the transmission line is read. A channel centerline and boundary zone description are constructed according to tower number, span, and conductor suspension point, generating a line safety parameter set. To ensure coordinate consistency, both types of parameters are mapped to a unified geographical reference. Tower foundation points and known baselines are used as control points. Back-projection is performed on the image coordinates to complete the coordinate system transformation, outputting a spatial feature matrix with mileage, lateral orientation, and height as columns. This matrix records the index of each sample and its corresponding tower segment, providing a primary key for subsequent partitioning and connectivity determination.

[0071] Based on the spatial feature matrix, a spatial grid is generated to produce a three-dimensional coordinate mapping table. During the grid division, the grid is segmented longitudinally with a fixed mileage step size, symmetrically divided laterally with the channel boundary as the zero point, and layered along the height according to the interval between the lowest point of the guide wire and the ground. A unique assignment rule is applied to samples falling on the grid boundaries to avoid duplicate counting. Each grid cell in the mapping table stores the sample count, centroid, and relative position marker with respect to the boundary zone, serving as input for distance calculation.

[0072] Based on the aforementioned three-dimensional coordinate mapping table, distance vector calculations are performed to obtain the coordinate mapping table. Specifically, three quantities are taken: the lateral distance from the centroid of each grid to the centerline, the vertical margin to the upper and lower boundaries, and the longitudinal spacing between adjacent grids along the line. Grid density is used as the reliability weight to form distance entries for connectivity analysis. The distance entries retain reference keys to the spatial feature matrix, facilitating the backtracking of original pixels and time points.

[0073] Based on the coordinate mapping table, a topological connectivity relationship is constructed to obtain a tree barrier spatial distribution model. This is achieved by creating edges between vertically adjacent grids with lateral differences less than a given threshold, and creating cross-layer edges between adjacent grids with continuous height margins. The two types of edges are then combined to obtain a set of connected components. For each connected component, its nearest boundary distance, coverage mileage, and height layer distribution are recorded as direct inputs for subsequent 3D reconstruction and protected area delineation. This model is then used in subsequent steps to generate spatial point clouds and voxel representations, and further used to extract the canopy projection range and growth trend.

[0074] In one embodiment of the AI-based tree obstacle threat level assessment method of this application, it may further include the following:

[0075] Step S301: Subdivide the tree barrier spatial distribution model into a grid according to the boundary constraints of the power facility protection area to obtain a spatial point cloud. Perform three-dimensional surface fitting on the spatial point cloud to generate a voxel model. Construct a spatial occupancy probability map based on the voxel model to obtain the region division result. Calculate the minimum bounding rectangle of the region division result to obtain the tree canopy projection range data.

[0076] Step S302: Perform time-series registration on the tree canopy projection range data to obtain a growth feature sequence, perform multi-dimensional mapping between the growth feature sequence and the tree canopy morphology parameter set to obtain a dynamic change matrix, and establish a growth rate model based on the dynamic change matrix to generate a growth trend set.

[0077] First, the aforementioned tree barrier spatial distribution model is read, and the three-dimensional space containing the connected components is subdivided into meshes according to three types of constraints: inner side, boundary zone, and outer side of the power facility protection area, outputting a spatial point cloud. During subdivision, the longitudinal distance, lateral distance, and height are used as the three-axis step sizes. Each connected component is sampled within an allowable range according to its boundary label, and restricted interpolation points are added for low-density areas to ensure continuous coverage for subsequent surface fitting. The spatial point cloud retains the source mesh and timestamp index for backtracking and noise reduction.

[0078] Based on the spatial point cloud, a 3D surface fitting is performed to generate a voxel model. During fitting, local surface estimation is performed on the point cloud of each connected component, and the estimation results are projected onto a fixed-resolution 3D voxel grid. The evidence count of voxel occupancy and the fitting residual are recorded. The occupancy confidence of voxels with out-of-bounds residuals is reduced to suppress isolated noise. The voxel model corresponds one-to-one with the connected components and carries a protected region label, providing a basis for probabilistic modeling.

[0079] Based on the voxel model, a spatial occupancy probability map is constructed to obtain the region partitioning results. During processing, the occupancy probability is calculated for each voxel, and corrections for inside-priority and boundary-band penalties are given in conjunction with the protected region label. Subsequently, maximum a posteriori (MAP) annotations for the three types of regions are performed at the voxel level, and the region partitioning volume is output. The partitioning volume is backfilled to the connected component attributes, and the voxel set and volume metric of each type of region are recorded, providing stable support for planar projection and temporal registration.

[0080] Based on the region division results, the minimum bounding rectangle is calculated to obtain the canopy projection range data. The steps involve projecting the voxel set of the inner side and boundary zone onto the ground plane, using a rotating caliper algorithm to obtain the minimum bounding rectangle of each connected contour segment, and outputting the projected area, major and minor axes, principal axis direction, and angle with the channel centerline, while retaining the mapping with timestamps and connected components. This data serves as the direct input for subsequent temporal registration.

[0081] Based on the canopy projection range data, temporal registration was performed to obtain the growth feature sequence. Registration was established using the similarity of the projection centroid distance and principal axis direction as dual constraints, creating intertemporal matching. When one-to-many or many-to-one relationships occurred, combinations continuous along the mileage were preferentially retained. The growth feature sequence recorded changes in projection area, principal axis length, lateral displacement, and region label over time, and referenced back to the original voxel evidence for subsequent discrepancy verification.

[0082] Based on the aforementioned growth characteristic sequence, it is mapped in multiple dimensions to the aforementioned canopy morphology parameter set to obtain a dynamic change matrix. During the mapping, the projected area and spatial height are aligned over time, and the area increment, lateral displacement increment, and height increment are calculated. These are then categorized and coded according to the protected area to form a unified time-series feature table. The dynamic change matrix uses the sample number as the primary key and lists the amount of change and direction marker for each time window.

[0083] Based on the dynamic change matrix, a growth rate model is established to generate a growth trend set. The model calculates the lateral approximation rate, vertical growth rate, and projection expansion rate for each sample, and uses window smoothing to suppress short-term fluctuations. When the lateral rate consistently points inward and is consistent with the region label change, it is classified as approximation type; when the vertical rate is dominant, it is classified as vertically dominant; when both are similar and their directions are stable, it is classified as mixed type. The growth trend set has the same primary key as the projection range data, allowing for direct reading in subsequent risk scoring and label classification.

[0084] In one embodiment of the AI-based tree obstacle threat level assessment method of this application, it may further include the following:

[0085] Step S401: Perform feature dimensionality reduction processing on the tree canopy projection range data to obtain the principal component feature set, perform projection transformation on the principal component feature set and the center line of the line channel to obtain the distance vector set, construct a K-nearest neighbor search tree based on the distance vector set to obtain the spatial approximate vector set, and calculate the danger level index set by Euclidean distance according to the boundary of the power facility protection area.

[0086] Step S402: Map the group of risk level indicators according to the safety level threshold to obtain a fuzzy evaluation set. Calculate the fuzzy evaluation set based on preset weight coefficients to obtain a comprehensive score vector. Normalize the comprehensive score vector with historical evaluation data to generate a risk score matrix.

[0087] First, the tree canopy projection range data and corresponding protected area labels are read, and fields such as projection area, major and minor axes, principal axis angles, and lateral displacement are standardized to form comparable input columns. Next, covariance decomposition is performed on the sample dimension, and principal component feature sets are obtained by truncating according to cumulative contribution. The variance proportion of the discarded components is recorded as residual information for subsequent stability reference during evaluation.

[0088] Based on the principal component feature set, a geometric mapping relationship is established with the centerline of the line corridor, and a projection transformation is performed to obtain a distance vector set. During the transformation, the parameterized mileage of the centerline is used as the vertical coordinate. The shortest lateral distance from each sample to the centerline, the mileage difference between adjacent samples along the line, and the relative direction sign of the principal axis are calculated, and the correspondence with the protected area label is preserved. Each record in the distance vector set is associated with the original sample primary key and timestamp to ensure traceability.

[0089] Based on the aforementioned distance vector set, a K-nearest neighbor search tree is constructed to obtain a spatial approximate vector set. The index is first layered according to the protected area label, and then a balanced tree structure is established within each layer using mileage and lateral distance. The principal axis direction sign is used as an additional key for pruning, limiting the search radius to the physically relevant range. The retrieved neighborhood samples are organized into fixed-length approximate vectors, carrying source density and time interval markers, which serve as input for distance calculation.

[0090] Based on the aforementioned spatial approximation vector set, Euclidean distance calculations are performed along the boundaries of the power facility protection zones to obtain a set of hazard level indicators. Three types of quantities are calculated and output: lateral distance to the channel boundary, distance along the line for neighboring connectivity, and vertical margin to the lowest point of the conductor. A residual-based stability penalty is added to samples within the boundary zone to suppress short-term noise. These quantities are unified to the same scale, forming indicator entries containing distance triples and penalty terms, which are archived along with the region labels as direct input for fuzzy evaluation.

[0091] Based on the aforementioned hazard level index set, a fuzzy evaluation set is obtained through membership degree mapping. The mapping rule defines three membership functions—low, medium, and high—based on safety level thresholds, which apply to the lateral, along-line, and vertical indicators, respectively. A stability penalty is also introduced to downgrade the membership degree of boundary zone samples. Each sample forms a membership degree vector, retaining the corresponding threshold source and time window for easy interpretation and traceability.

[0092] Based on the fuzzy evaluation set, a comprehensive score vector is calculated according to preset weight coefficients. Weights are assigned hierarchically based on the reliability of the indicator source and regional labels. First, intra-dimensional synthesis is performed within the same indicator, then inter-dimensional weighting is applied across the three indicators to obtain a single-row score for each sample. The score entries reference principal component residual information for subsequent normalization scaling.

[0093] Based on the comprehensive scoring vector, it is combined with historical evaluation data for normalization to generate a risk scoring matrix. Normalization employs piecewise linear stretching and robust center alignment to ensure comparability across different timeframes and seasonal conditions. Simultaneously, samples with data gaps are given conservative scores through neighborhood interpolation, and uncertain regions are separately labeled. The final risk scoring matrix uses sample primary keys as rows and risk dimensions as columns, which is used for subsequent grading mapping to generate tree barrier level labels and is directly read by the dynamic evaluation model during the training phase.

[0094] In one embodiment of the AI-based tree obstacle threat level assessment method of this application, it may further include the following:

[0095] Step S501: Divide the risk scoring matrix into risk level clusters according to hierarchical clustering rules, generate an initial label set by label encoding of the risk level clusters based on the safety distance constraint, verify the association between the initial label set and historical control records to obtain a tree barrier level label set, and construct a training sample library based on the tree barrier level label set.

[0096] Step S502: Perform feature fusion on the label data in the training sample library and the growth trend set to obtain a training feature matrix. Input the training feature matrix into a deep neural network for supervised learning to obtain a prediction model. Perform cross-validation and parameter optimization on the prediction model to generate a dynamic evaluation model.

[0097] First, the aforementioned risk scoring matrix is ​​read, and a standardized input block is established with samples as rows and risk dimensions as columns. Anomalies are then weighted and masked based on source reliability and residual labels. Hierarchical partitioning is performed on this input block according to hierarchical clustering rules. Aggregation is first performed within the same protection area using Euclidean metrics, and then a small penalty is introduced for merging across regions, outputting risk level clusters. Each cluster is recorded with a member list, center vector, and boundary sample index for subsequent label encoding and consistency verification.

[0098] Based on the aforementioned risk level clusters, and according to the rigid boundaries and boundary buffer clauses of the safety distance for power facilities, an initial label set is generated through label encoding. Specifically, the three dimensions of the center vector—lateral, along-line, and vertical—are compared one by one with the constraint interval: those completely falling within the safety interval are labeled as low risk, those with single-dimensional boundary violations are labeled as medium risk, and those with multi-dimensional boundary violations are labeled as high risk; uncertain markers are retained for boundary samples, and trigger dimensions are recorded. The initial label set is expanded to the sample level at the cluster level, while retaining intra-cluster consistency scores for backtracking.

[0099] Based on the initial label set, historical control records are loaded and aligned by mileage and time window for correlation verification. For samples that have undergone cleanup or re-inspection and whose record results are stable, the consistency between their labels and actual control effects is checked; if systematic deviations exist, the label decisions for the boundary samples of the corresponding cluster are adjusted, and the reasons for the adjustment are written in the annotation. After verification, a tree barrier level label set is obtained, containing two categories: definite labels and uncertain areas, along with source evidence pointers.

[0100] Based on the aforementioned tree barrier level label set, a training sample library is constructed. The training sample library uses sample primary keys as indexes, linking risk scoring matrix entries, labels, cluster information, and historical evidence to form a structured record that can be directly fed into the learning model. To avoid information leakage, adjacent time slices within the same mileage segment are segmented by scenario to ensure that the same event chain is not simultaneously included in training and validation.

[0101] Based on the training sample library, the labeled data is fused with the aforementioned growth trend set to obtain a training feature matrix. During fusion, the risk dimension score is concatenated column-wise with the trend direction and trend rate, and time location encoding and region label encoding are introduced. Samples in the uncertainty area are assigned reduced sample weights to prevent overfitting to boundary samples during the training phase. The training feature matrix retains the reference key to the projection range data to support subsequent error analysis.

[0102] Based on the trained feature matrix, a deep neural network is input for supervised learning to obtain a prediction model. This network, named the Tree Barrier Dynamic Discrimination Network, includes a temporal feature extraction branch and a static risk branch, which are fused at a fully connected layer to output a rank probability distribution. During training, weighted cross-entropy is used as the objective, sample weights are derived from label determinism and source reliability, and a trend consistency constraint is set as a regularization term to suppress short-term direction reversals.

[0103] Based on the prediction model, cross-validation and parameter optimization are performed to generate a dynamic evaluation model. Cross-validation is performed by segmenting the line to ensure spatial independence between segments. Parameter optimization employs a joint grid search with learning rate and weight decay, using the rank accuracy on the validation set and the violation rate of the uncertainty zone constraint as dual objectives to select the optimal parameters. The final dynamic evaluation model, along with the threshold reference and feature specifications, is solidified into a deployable entity for direct loading in subsequent real-time monitoring steps, and is referenced in a consistent manner with the aforementioned label mapping relationship.

[0104] In one embodiment of the AI-based tree obstacle threat level assessment method of this application, it may further include the following:

[0105] Step S601: Convert the dynamic evaluation model into a lightweight inference engine to obtain a prediction computing unit, perform model quantization and pruning optimization on the prediction computing unit to obtain a real-time inference model, deploy the real-time inference model to the edge computing node of the monitoring system, and build a data preprocessing pipeline based on the edge computing node to obtain a real-time analysis unit.

[0106] Step S602: Receive tree obstacle monitoring data through the real-time analysis unit and extract features to obtain a monitoring feature vector. Input the monitoring feature vector into the real-time inference model to predict the threat level and obtain a predicted probability distribution. Determine the threshold of the predicted probability distribution according to the early warning rules to generate a threat early warning result.

[0107] First, the dynamic evaluation model and label mapping relationship are read, and the network structure, feature specifications, and threshold references are imported into a conversion tool to generate a prediction computation unit for edge devices. During the conversion, operators are fused and shape inference is performed, the temporal length and feature column order of the input tensor are fixed, and the direction and rate fields of the growth trend set are set as optional input channels so as to fall back to the static risk branch when there is no trend data. The prediction computation unit also exports a field validation checklist for automatic alignment during the deployment phase.

[0108] Based on the prediction calculation unit, model quantization and pruning are performed to obtain the real-time inference model. The quantization stage is calibrated based on representative samples, standardizing the fixed-point format of activation and weights, and recording the dynamic range to support saturation protection. The pruning stage prunes low-contribution branches according to channel importance statistics, retaining key residual connections for time-series branches to avoid excessive loss of short-window inputs. After these two steps, a closed-loop verification is performed to confirm that the output level distribution is consistent with the original model within acceptable deviations.

[0109] Based on the real-time inference model, edge deployment is completed, and the model is loaded onto the edge computing nodes of the monitoring system. During deployment, a hardware acceleration backend is bound, batch size and concurrent queue are set, and asynchronous I / O is enabled to connect to multi-source sensors. To ensure spatiotemporal consistency, a unified clock source is enabled within the node, and image frames, point cloud frames, and trajectory positioning are aligned according to the observation and reporting time. Frames exceeding the late threshold are backfilled using a compensation strategy to prevent empty segments from participating in inference.

[0110] Based on the edge computing nodes, a data preprocessing pipeline is constructed to obtain the real-time analysis unit. The pipeline is organized into "access, verification, geometric alignment, feature generation, and cache write-back": the access stage parses data packets and metadata; the verification stage filters out missing columns and abnormal frames according to the field verification list; the geometric alignment stage completes projection transformation and coordinate registration; the feature generation stage outputs lateral distance, projection principal axis, spatial height, stability markers, and optional trend direction and rate; the cache write-back stage associates features with the original evidence index for subsequent alarm backtracking.

[0111] Based on the real-time analysis unit, tree obstacle monitoring data is received and monitoring feature vectors are generated. The features are organized as fixed-column tensors, with the column order consistent with the prediction calculation unit. Missing trend columns are placed using zero masks, and source reliability weights are added as sample weight references during inference. The generated features are then batch-entered into the real-time inference model for forward computation.

[0112] Based on the real-time inference model, the predicted probability distribution and trend consistency score of the threat level are output. The inference result, along with the sample weights, is sent to the threshold determination module on the edge side. The module compares each rule according to the warning rules, including three conditions: minimum safe distance threshold, boundary buffer, and consistency constraint. If the highest level of the probability distribution exceeds the corresponding threshold and the consistency score meets the directional constraint, a threat warning result is generated, recording the trigger dimension, evidence index, and time window.

[0113] Based on the threat warning results, duplicate suppression and merging are performed. Alarms of the same type within a continuous time window of the same mileage segment are aggregated, outputting a single continuous event, and the interval between the first trigger and the most recent trigger is calculated for downstream scheduling. Alarms that are adjacent across segments and caused by the same connected component are merged using a channel proximity priority rule to avoid duplicate dispatch.

[0114] Based on the merged early warning entries, on-site executable instructions are encapsulated. The entries are written to the dispatch queue of the corresponding edge node, including the target mileage segment, area label, suggested handling level, and verification collection requirements. Simultaneously, the entry is written back to the central monitoring database to maintain consistent referencing with the risk scoring matrix and label mapping. The encapsulation result serves as the entry point for subsequent hierarchical management, allowing the scheduling module to directly read and issue commands.

[0115] In one embodiment of the AI-based tree obstacle threat level assessment method of this application, it may further include the following:

[0116] Step S701: Align the threat warning results with historical inspection records according to the spatiotemporal index to obtain a related dataset. Perform time-series correlation analysis on the related dataset to obtain risk evolution characteristics. Construct a warning rule base based on the risk evolution characteristics to obtain a warning strategy model. Perform rule matching between the warning strategy model and real-time warning data to generate warning instructions.

[0117] Step S702: Establish a control priority queue based on the threat level of the warning instruction to obtain a task scheduling table. Divide the task scheduling table into partitions according to the work resource constraints to obtain a work assignment scheme. Implement differentiated control measures for tree obstacles on transmission lines based on the work assignment scheme and record the control results.

[0118] First, threat warning results and historical inspection records are linked using a joint key based on three fields: channel mileage, tower number, and timestamp. This completes spatiotemporal index alignment, outputting a correlated dataset. During alignment, entries with location drift are relocated using the conductor direction as a constraint, and placeholder markers are inserted for missing time periods to ensure a continuous sequence of the same object on the timeline. The correlated dataset retains the source device, sampling window, and evidence index for easy backtracking later.

[0119] Based on the aforementioned associated dataset, time-series correlation analysis is conducted to obtain risk evolution characteristics. The analysis process extracts graded trajectories within the same mileage segment, cross-seasonal recurring locations, alarm duration length, and intermittent periods, and calculates synchronization indicators with meteorological records. For objects with multi-source evidence, a consistency score is added to identify stable risks and occasional disturbances. In the risk evolution characteristics, each entry points back to the corresponding early warning result and inspection record key, forming an interpretable evolution path.

[0120] Based on the aforementioned risk evolution characteristics, an early warning rule base is constructed, and an early warning strategy model is generated. The rule base is divided into three categories: continuously weighted rules, repeatedly triggered rules, and critical approximation rules; each type of rule clearly defines the applicable level threshold conditions, minimum duration window, and merging strategy. The early warning strategy model is organized in the form of a decision table, loading the corresponding rule set according to line segments, and recording the binding relationship with the protection area label to constrain the scope of rule effectiveness.

[0121] Based on the aforementioned early warning strategy model, rule matching is performed with real-time early warning data to generate early warning instructions. The matching process evaluates each rule sequentially using the object's primary key. When the triggering condition of a rule matches the real-time data, an instruction entry containing the target mileage segment, suggested handling level, review requirements, and time requirements is output. If multiple rules are matched simultaneously, a decision is made according to the priority stack, and the source of the conflict and the reason for the decision are written. The early warning instruction writes back its source rule number and evidence index to ensure closed-loop tracking.

[0122] Based on the aforementioned early warning instructions, a control priority queue is established according to the threat level, resulting in a task scheduling table. During scheduling, emergency response, time-limited re-inspection, and routine observation are respectively placed into high, medium, and low queues, and adjacent objects within the same segment are merged to form continuous work areas, reducing repeated start-ups and shutdowns. The task scheduling table records the target location, work window, required personnel skills, and safety measures for each task, serving as a structured entry point before dispatch.

[0123] Based on the task scheduling table, the task allocation scheme is obtained by partitioning the work area according to the resource constraints. Resource constraints include vehicle accessibility, terrain limitations, power outage planning window, shift load, and tool configuration. After partitioning, candidate and backup shifts are assigned to each work area, and arrival order and routing suggestions are given; for segments with consecutive objects across areas, synchronization points are inserted to avoid duplicate blocking.

[0124] Based on the aforementioned task assignment plan, differentiated management and control measures are implemented for tree obstructions along transmission lines, and the results are recorded. Emergency response tasks are issued with dual instructions for clearing and on-site verification; medium-risk tasks are issued with a time-limited re-inspection and image transmission requirements; low-risk tasks are assigned observation and a new observation window is set. Upon completion of the task, the location, images, and task description are collected, updated in the task record, and a one-to-one correspondence is established with the original warning instructions.

[0125] Based on the control results, execution verification and closed-loop write-back are performed. Verification is based on comparing the achieved status with the level results of the on-site review images and the re-inferenced levels; if there is a significant deviation from the warning level, the applicable scope of the rule is marked in the strategy model and the priority is adjusted. Finally, the achieved status, the reasons for non-achievement, and the suggested improvements are written back to the risk evolution characteristics, which are used as the entry point for rule revision and model retraining in subsequent cycles, ensuring the continuity and reproducibility of the process from alarm generation to on-site handling.

[0126] To effectively address the shortcomings of traditional technologies in tree obstacle modeling, threat assessment, and early warning and control, and to provide technical support for the safety management of transmission lines, this application provides an embodiment of an AI-based tree obstacle threat level assessment device for implementing all or part of the aforementioned AI-based tree obstacle threat level assessment method. See [link to embodiment]. Figure 2 The AI-based tree obstacle threat level assessment device specifically includes the following components:

[0127] The tree obstacle identification module 10 is used to acquire a set of tree crown morphology parameters and a set of line safety parameters based on high-definition images of transmission lines, construct a coordinate mapping table according to the safety distance regulations for power facilities, generate a tree obstacle spatial distribution model through topological analysis of the coordinate mapping table, perform three-dimensional reconstruction of the tree obstacle spatial distribution model to obtain the protection area division result, extract tree crown projection range data based on the protection area division result, and perform correlation analysis between the tree crown projection range data and the tree crown morphology parameter set to generate a growth trend set.

[0128] Tree barrier analysis module 20 is used to construct a spatial approximate vector set from the tree canopy projection range data, perform distance calculation on the spatial approximate vector set to obtain a risk level index group, perform fuzzy comprehensive evaluation on the risk level index group to generate a risk score matrix, perform hierarchical mapping based on the risk score matrix to obtain a tree barrier level label set, and perform collaborative training of the tree barrier level label set and the growth trend set to obtain a dynamic evaluation model.

[0129] The risk warning module 30 is used to deploy the dynamic assessment model in the real-time monitoring system, receive tree obstacle monitoring data and make predictions to obtain threat warning results, correlate the threat warning results with historical inspection records to generate warning instructions, and perform hierarchical management and control of tree obstacles on transmission lines according to the warning instructions.

[0130] As described above, the AI-based tree obstacle threat level assessment device provided in this application can achieve accurate tree obstacle modeling through spatial analysis and 3D reconstruction. An assessment mechanism is constructed, combining comprehensive evaluation and dynamic assessment to establish a reliable threat early warning strategy. Control optimization is introduced, ensuring the continuous safety of the transmission line through early warning prediction and hierarchical control. This method effectively solves the shortcomings of traditional technologies in tree obstacle modeling, threat assessment, and early warning control, providing technical support for the safety management of transmission lines.

[0131] This invention also provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the AI-based tree barrier threat level assessment method.

[0132] This invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the aforementioned AI-based tree obstacle threat level assessment method.

[0133] This invention also provides a computer program product, which includes a computer program that, when executed by a processor, implements the aforementioned AI-based tree obstacle threat level assessment method.

[0134] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0135] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0136] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0137] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0138] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. An AI-based method for assessing tree obstacle threat levels, characterized in that, The method includes: Based on high-definition images of transmission lines, a tree canopy morphology parameter set and a line safety parameter set are acquired. A coordinate mapping table is constructed according to the power facility safety distance regulations. This coordinate mapping table is then used to generate a tree barrier spatial distribution model through topological analysis. The tree barrier spatial distribution model is then reconstructed in three dimensions to obtain the protection area division results. Based on the protection area division results, tree canopy projection range data is extracted. The tree canopy projection range data is then correlated with the tree canopy morphology parameter set to generate a growth trend set. This process includes: subdividing the tree barrier spatial distribution model into a grid according to the power facility protection area boundary constraints to obtain a spatial point cloud; performing three-dimensional surface fitting on the spatial point cloud to generate a voxel model; constructing a spatial occupancy probability map based on the voxel model to obtain the area division results; calculating the minimum bounding rectangle of the area division results to obtain the tree canopy projection range data; performing temporal registration on the tree canopy projection range data to obtain a growth feature sequence; mapping the growth feature sequence with the tree canopy morphology parameter set in multiple dimensions to obtain a dynamic change matrix; and establishing a growth rate model based on the dynamic change matrix to generate a growth trend set. A spatial approximate vector set is constructed from the tree canopy projection range data. Distance calculation is performed on the spatial approximate vector set to obtain a group of risk level indicators. Fuzzy comprehensive evaluation is performed on the group of risk level indicators to generate a risk score matrix. A tree barrier level label set is obtained by hierarchical mapping based on the risk score matrix. The tree barrier level label set and the growth trend set are co-trained to obtain a dynamic evaluation model. The dynamic assessment model is deployed in a real-time monitoring system to receive tree obstacle monitoring data and make predictions to obtain threat warning results. The threat warning results are correlated with historical inspection records to generate warning instructions. Based on the warning instructions, the tree obstacles on the transmission line are managed in a hierarchical manner.

2. The AI-based tree barrier threat level assessment method according to claim 1, characterized in that, The method involves acquiring a set of tree canopy morphology parameters and a set of line safety parameters based on high-definition images of transmission lines, constructing a coordinate mapping table according to the safety distance regulations for power facilities, and generating a spatial distribution model of tree barriers through topological analysis of the coordinate mapping table, including: Based on high-definition images of transmission lines, the vertical projection area and spatial height data of tree canopies are extracted. The vertical projection area and spatial height data of tree canopies are normalized according to the safety distance regulations for power facilities to obtain a tree canopy morphology parameter set. A line safety parameter set is generated according to the network topology structure from the substation to the transmission line. The tree canopy morphology parameter set and the line safety parameter set are transformed into coordinate systems to obtain a spatial feature matrix. The spatial feature matrix is ​​divided into spatial grids to generate a three-dimensional coordinate mapping table. The distance vector between the tree obstacle location information in the three-dimensional coordinate mapping table and the boundary of the route channel is calculated to obtain a coordinate mapping table. Based on the coordinate mapping table, a topological connectivity relationship is constructed to obtain a tree obstacle spatial distribution model.

3. The AI-based tree obstacle threat level assessment method according to claim 1, characterized in that, The step of constructing a spatial approximate vector set from the tree canopy projection range data, calculating the distance of the spatial approximate vector set to obtain a risk level index group, and performing a fuzzy comprehensive evaluation on the risk level index group to generate a risk scoring matrix includes: The tree canopy projection range data is subjected to feature dimensionality reduction processing to obtain a principal component feature set. The principal component feature set is then projected onto the centerline of the power line channel to obtain a distance vector set. A K-nearest neighbor search tree is constructed based on the distance vector set to obtain a spatial approximate vector set. The spatial approximate vector set is then calculated using Euclidean distance according to the boundary of the power facility protection area to obtain a hazard level index set. The risk level index group is mapped to a fuzzy evaluation set according to the safety level threshold. The fuzzy evaluation set is weighted based on a preset weight coefficient to obtain a comprehensive score vector. The comprehensive score vector is normalized with historical evaluation data to generate a risk score matrix.

4. The AI-based tree barrier threat level assessment method according to claim 1, characterized in that, The process of obtaining a tree barrier level label set by performing a hierarchical mapping based on the risk scoring matrix, and then co-training the tree barrier level label set with the growth trend set to obtain a dynamic evaluation model, includes: The risk scoring matrix is ​​hierarchically divided according to hierarchical clustering rules to obtain risk level clusters. The risk level clusters are then labeled and encoded based on safe distance constraints to generate an initial label set. The initial label set is then correlated and verified with historical control records to obtain a tree barrier level label set. A training sample library is then constructed based on the tree barrier level label set. The label data in the training sample library and the growth trend set are fused to obtain a training feature matrix. The training feature matrix is ​​then input into a deep neural network for supervised learning to obtain a prediction model. The prediction model is then cross-validated and its parameters are optimized to generate a dynamic evaluation model.

5. The AI-based tree obstacle threat level assessment method according to claim 1, characterized in that, The step of deploying the dynamic assessment model in a real-time monitoring system, receiving tree obstacle monitoring data, and making predictions to obtain threat warning results includes: The dynamic evaluation model is converted into a lightweight inference engine to obtain a prediction computing unit. The prediction computing unit is then subjected to model quantization and pruning optimization to obtain a real-time inference model. The real-time inference model is deployed to the edge computing node of the monitoring system. A data preprocessing pipeline is then constructed based on the edge computing node to obtain a real-time analysis unit. The real-time analysis unit receives tree obstacle monitoring data and extracts features to obtain a monitoring feature vector. The monitoring feature vector is then input into the real-time inference model to predict the threat level and obtain a predicted probability distribution. The predicted probability distribution is then used to determine a threshold according to the early warning rules to generate a threat early warning result.

6. The AI-based tree obstacle threat level assessment method according to claim 1, characterized in that, The step of correlating and analyzing the threat warning results with historical inspection records to generate warning instructions, and then implementing graded control of tree obstructions on transmission lines based on the warning instructions, includes: The threat warning results and historical inspection records are aligned with the spatiotemporal index to obtain a related dataset. The temporal correlation analysis of the related dataset is performed to obtain risk evolution characteristics. Based on the risk evolution characteristics, a warning rule base is constructed to obtain a warning strategy model. The warning strategy model is matched with real-time warning data to generate warning instructions. Based on the threat level of the warning instruction, a control priority queue is established to obtain a task scheduling table. The task scheduling table is partitioned according to the operation resource constraints to obtain an operation assignment scheme. Based on the operation assignment scheme, differentiated control measures are implemented for tree obstacles on transmission lines and the control results are recorded.

7. An AI-based tree obstacle threat level assessment device, characterized in that, The device includes: The tree obstacle identification module is used to acquire tree canopy morphology parameter sets and line safety parameter sets based on high-definition images of transmission lines. It constructs a coordinate mapping table according to the power facility safety distance regulations, generates a tree obstacle spatial distribution model through topological analysis of the coordinate mapping table, performs 3D reconstruction of the tree obstacle spatial distribution model to obtain protection area division results, extracts tree canopy projection range data based on the protection area division results, and performs correlation analysis between the tree canopy projection range data and the tree canopy morphology parameter set to generate a growth trend set. This includes: subdividing the tree obstacle spatial distribution model into a grid according to the power facility protection area boundary constraints to obtain a spatial point cloud; performing 3D surface fitting on the spatial point cloud to generate a voxel model; constructing a space occupancy probability map based on the voxel model to obtain the area division results; calculating the minimum bounding rectangle of the area division results to obtain tree canopy projection range data; performing temporal registration on the tree canopy projection range data to obtain a growth feature sequence; mapping the growth feature sequence with the tree canopy morphology parameter set in multiple dimensions to obtain a dynamic change matrix; and establishing a growth rate model based on the dynamic change matrix to generate a growth trend set. The tree barrier analysis module is used to construct a spatial approximate vector set from the tree canopy projection range data, perform distance calculation on the spatial approximate vector set to obtain a risk level index group, perform fuzzy comprehensive evaluation on the risk level index group to generate a risk score matrix, perform hierarchical mapping based on the risk score matrix to obtain a tree barrier level label set, and perform collaborative training of the tree barrier level label set and the growth trend set to obtain a dynamic evaluation model. The risk warning module is used to deploy the dynamic assessment model on the real-time monitoring system, receive tree obstacle monitoring data and make predictions to obtain threat warning results, correlate the threat warning results with historical inspection records to generate warning instructions, and perform hierarchical management and control of tree obstacles on transmission lines according to the warning instructions.

Images