A control network construction method based on dynamic optimization of unmanned aerial vehicle reconnaissance intelligent agents

Abstract

This application relates to the field of navigation and control technology, and in particular to a control network construction method based on dynamic optimization by an AI agent using a UAV. The method involves using a UAV equipped with a panoramic camera and a GNSS measurement module to collect control point data in real time. This data is then fed back to the AI ​​agent, which dynamically corrects the control point suitability factors and iteratively optimizes the control network points. The method includes a multi-factor dynamic evaluation and screening model for control points, and a mechanism for dynamically adjusting factor weights based on project objectives. For scenarios such as high mountains and dense forests, the method utilizes the UAV to identify temporary obstructions or terrain obstacles, and the AI ​​agent automatically adds candidate points to areas with accuracy gaps. This invention significantly improves the accuracy and reliability of the control network, substantially increases operational efficiency, shortens the control network design cycle, enhances adaptability to complex environments, and, through multi-source data fusion, can accurately address interference such as occlusion and multipath effects.

Description

Technical Field

[0001] This application relates to the field of navigation and control technology, and in particular to a method for constructing a control network based on dynamic optimization of UAV reconnaissance agents. Background Technology

[0002] Global Navigation Satellite System (GNSS) technology can efficiently acquire large-scale, high-precision spatial coordinate data, playing an increasingly crucial role in fields such as geodesy, engineering construction, geological monitoring, and smart cities. Deeply integrated into modern surveying systems, it breaks through the limitations of traditional control surveying, which relies on contact-based and step-by-step measurements such as triangulation networks and traverse networks, achieving automated, all-weather, and high-efficiency field observations. GNSS technology rapidly acquires the three-dimensional coordinate information of control points, providing core data support for applications such as regional mapping benchmark establishment, engineering deformation monitoring, and geographic information updates. The design quality of the GNSS control network directly determines the accuracy and reliability of subsequent surveying results, while optimized design is a key step in balancing accuracy, cost, and efficiency while meeting project requirements. Therefore, scientifically optimizing the structure, benchmark, and observation scheme of the control network has become an important and fundamental task.

[0003] Currently, the main methods for optimizing GNSS control networks include: network structure optimization methods, which enhance geometric strength by adjusting the distribution and connection relationships of control points. The core idea is to improve the network's error resistance by uniformly distributing control points and adding closed loops and cross connections, essentially adding more robust nodes and connecting lines to a network, significantly improving its overall stability; benchmark selection optimization methods, used to solve the coordinate and scale benchmark problems of the control network, avoiding excessive propagation of benchmark errors by rationally selecting the number and distribution of starting points; observation scheme optimization methods, which reduce the impact of satellite signal interference and environmental errors by selecting the optimal observation period and adjusting the observation duration and sampling rate; and numerical optimization algorithms, such as genetic algorithms and simulated annealing algorithms, which solve for the optimal solution under multiple constraints by establishing mathematical models. These methods provide a solid theoretical foundation and practical guidance for the design and optimization of GNSS control networks.

[0004] Existing GNSS control network design and optimization methods provide theoretical and practical support for the scientific deployment of control networks. However, in complex areas such as those with large elevation differences and high mountain terrain, manual on-site reconnaissance often leaves many areas inaccessible. The straight-line distance between control points can be 3-5 km, requiring several hours of driving around mountains, severely impacting survey efficiency. Currently, GNSS control network design suffers from the following main problems: First, low efficiency of manual surveys: single-point surveys in mountainous areas require ≥2 hours, with an accident rate exceeding the industry average by 3 times. Second, insufficient reliability of initial selection: traditional pre-selection relies on single indicators such as DEM slope, failing to incorporate dynamic factors such as geological stability and signal obstruction probability. Third, limitations of static optimization: network adjustments are disconnected from field data; network design relies on preliminary terrain data and cannot be dynamically adjusted according to the actual environment, resulting in 85% of projects requiring rework due to problems discovered during later construction. Existing "GNSS + total station layered control networks" still require manual pre-setting of benchmark points, and "adaptive control networks" rely on fixed station laser scanning, making it difficult to achieve fully automated dynamic optimization. Existing literature suggests that while network structure optimization can enhance geometric strength by adjusting the distribution and connection of control points, in steep mountainous terrain, accessibility limitations often force control point placement in areas with signal obstruction, making it difficult to balance point feasibility and geometric strength, and easily creating areas of weak local accuracy. Benchmark selection optimization, when dealing with coordinate and scale benchmarks, relies excessively on historical data to judge stability due to active crustal activity in high-altitude areas and the difficulty of frequent remeasurement of starting points in remote locations; sudden geological activity can easily lead to the propagation of errors throughout the network. Observation scheme optimization depends on satellite ephemeris and environmental parameters, but the variable weather and complex electromagnetic interference due to terrain reflection in high-altitude terrain, coupled with the difficulty of timely equipment adjustments in remote areas, make it difficult to cope with sudden interference, leading to a decline in data quality. When numerical optimization algorithms such as genetic algorithms process such large-scale control networks, terrain factors increase the dimensionality of variables, easily resulting in slow convergence and difficulty in escaping local optima. Although some scholars have improved performance to some extent by introducing terrain-adaptive weights and combining real-time monitoring to correct benchmarks, further research is needed on its adaptability to extreme conditions. Therefore, there is an urgent need to deeply improve and optimize the methods to balance computational stability with engineering-acceptable efficiency. Summary of the Invention

[0005] To achieve the above objectives, this application provides the following technical solution: According to a first aspect of the present invention, the present invention claims a method for constructing a control network based on dynamic optimization of an unmanned aerial vehicle (UAV) reconnaissance agent, comprising the following steps: S1: Construct a control network intelligent agent. The control network intelligent agent adopts a four-layer architecture of data layer, feature layer, model layer and application layer. It performs multi-source data fusion and analysis, candidate point adaptability scoring and preliminary selection result optimization on candidate control points in the survey area, and outputs a preliminary selection list of control points. S2: Based on the preliminary control point list, plan the UAV reconnaissance route, and use the UAV equipped with a panoramic high-definition camera, GNSS high-precision receiver and laser rangefinder to conduct on-site reconnaissance of each preliminary control point, and collect GNSS signal measurement data and environmental data. S3: Compare the measured data collected by the UAV with the candidate point information initially generated by the intelligent agent, and correct the satellite signal obstruction rate, PDOP value, terrain stability, spatial distribution rationality and interference signal source optimization factor; S4: Input the corrected optimization factors into the control network agent to perform dynamic optimization of the control network, including secondary evaluation to remove unqualified points, adding new points to optimize the distribution of the control network, and optimizing the connection relationship between points through the minimum spanning tree algorithm. S5: Output the optimal control network results, including a list of control points, a visualization scheme, and a field operation guide.

[0006] Furthermore, step S1 also includes: The data layer is used for inputting and preprocessing multi-source data, including terrain data, redline data, occlusion data, satellite signal prediction data, and auxiliary data, and performs coordinate system 1, ground filtering, and rasterization processing. The feature layer is used to extract key features from the preprocessed data, including redline avoidance, terrain stability, satellite signal quality, spatial distribution, road accessibility, and interference. The model layer employs a basic model combining a lightweight convolutional neural network and a gradient boosting tree, and integrates a rule engine for hard constraint verification. The application layer is used to output a list of coordinates of the initial control points and a visual map, and provides an iterative mechanism to fine-tune the model.

[0007] Furthermore, the preprocessing steps of the data layer include: S121: Convert all input data to the 2000 National Geodetic Coordinate System to ensure spatial consistency of data; S122: Perform ground filtering on the LiDAR point cloud to separate ground points from vegetation or building points, and calculate the sky openness of each candidate point based on the filtered point cloud, wherein the sky openness is defined as the unobstructed zenith angle range. S123: Rasterize the satellite signal prediction data to generate a signal quality heatmap in grid cells, where each grid cell contains PDOP value, number of satellites and obstruction rate attributes. The multi-source data includes: High-resolution DEM data for slope analysis; The red line data, which consists of ecological protection red lines, permanent basic farmland, historical and cultural relic protection lines, and geological disaster risk lines, should be input in vector format. DSM data or LiDAR point clouds are used to extract tree and building height occlusion factors. Satellite visibility simulation data based on the latitude, longitude and topography of the survey area, including PDOP values ​​and satellite obstruction rates at different times; Auxiliary data consisting of traffic road data, geological disaster risk lines, high-voltage power transmission lines, and reservoir water surface distribution.

[0008] Furthermore, the key feature extraction of the feature layer includes: S131: Extract red line avoidance features, and determine whether the coordinates of candidate points fall within the range of ecological protection red lines, permanent basic farmland, historical and cultural relics protection lines or geological disaster risk lines by superimposing the vector data of the survey area. If they are within the range, they are marked as red line violations. S132: Extract terrain stability features and calculate the slope and aspect of each candidate point based on DEM data, where the slope is in degrees and the aspect is in azimuth. S133: Extract satellite signal quality features. Obtain the number of satellites, PDOP value and occlusion rate corresponding to candidate points from the rasterized signal quality heat map. The occlusion rate is calculated by sky openness. S134: Extract spatial distribution features, calculate the distance between control points and the regional point density based on the coordinates of candidate points, where the point density is expressed as the number of points per square kilometer; S135: Extract road accessibility features, calculate the straight-line distance from candidate points to the nearest road based on traffic road vector data, and assign a difficulty score based on terrain complexity; S136: Extract interference features and calculate the straight-line distance between candidate points and high-voltage transmission lines and reservoir surfaces. If the distance is less than the threshold, it is marked as interference risk.

[0009] Furthermore, the specific implementation of the model layer includes: A lightweight convolutional neural network is used to process rasterized signal quality heatmaps and terrain data to extract spatial features; Gradient boosting trees are used to process structured features, including slope, distance, and density, and output the point scores of candidate points. The rule engine performs hard constraint verification on the points output by the model, forcibly removing points with a PDOP value greater than 6, a number of satellites less than 4, or located within the red line range, and conditionally removing points with a spacing of less than 500 meters and redundant signal quality.

[0010] Furthermore, the specific steps of the on-site reconnaissance by UAV in step S2 include: Step S21: Plan the UAV flight path based on the preliminary control point list, ensuring that the flight path covers all candidate points with a score of 70 or higher; Step S22: The UAV flies to each candidate point in sequence, hovers at a height of 15-2 meters, simulates the GNSS receiver environment, hovers for 10-15 minutes to collect GNSS data, including 10 sets of single-point measurements, smoothing each set 10 times, and recording the number of satellites, PDOP value, carrier-to-noise ratio and multipath error; Step S23: Use a 360-degree panoramic high-definition camera to take panoramic photos of each point. The gimbal rotates one frame every 30 degrees, and the resolution is no less than 12 million pixels. This is used to verify vegetation obstruction, terrain changes and sources of interference. Step S24: Obtain the three-dimensional coordinates of the UAV through the GNSS receiver, and combine them with the altitude measured by the laser rangefinder to geometrically calculate the three-dimensional coordinates of the preset control point; Step S25: Transmit the panoramic photos, GNSS data, and coordinate information to the computer processing software in real time.

[0011] Furthermore, step S3 also includes: For satellite signal obstruction rate, compare the actual obstruction rate of the UAV with the predicted value of the agent. If the actual value is more than 15% higher than the predicted value, the score of this factor will be directly reduced. For PDOP values, if the measured PDOP value of the UAV is greater than 6 and the predicted value is less than or equal to 6, it is marked as a signal risk point and the PDOP factor is corrected. For terrain stability, the slope value calculated by DEM is corrected based on the detailed slope and foundation condition captured by drone. If loose foundation or hidden steep slope is found, the score is adjusted. To ensure the rationality of spatial distribution, the theoretical point spacing was adjusted based on the actual terrain obstacles captured by drone aerial photography. For sources of interference, the actual distance to high-voltage lines or reservoirs is determined based on high-definition images from drones. If the images show an unpredictable source of interference, the distance value is adjusted.

[0012] Furthermore, step S4 also includes: Step S41: Secondary evaluation and elimination. The candidate point fit is recalculated using the corrected factors. Points with a measured PDOP value greater than 6 and lasting for more than 5 minutes, actual occlusion rate greater than 50%, or less than 50 meters from the high-voltage line or less than 200 meters from the reservoir surface are eliminated. Step S42: Gap filling. For the accuracy gap area that appears after elimination, add new candidate points within a 1-2 kilometer radius, prioritizing areas that are verified by UAV as having low occlusion and a PDOP value of less than or equal to 6. Step S43: Distribution optimization. The minimum spanning tree algorithm is used to optimize the connection relationship between the retained points to ensure that there are no obstacles confirmed by the UAV between adjacent points and that the average spacing meets the accuracy requirements of the control network. Step S44: Multi-objective balance verification to ensure that the final control network has a score of 70 or higher, the maximum PDOP of the entire network is less than or equal to 6, the total number of points decreases by no more than 15%, and at least 30% of the points can be used as benchmark points.

[0013] Furthermore, the output of the optimal control network in step S5 includes: The control point list lists the coordinates, measured PDOP value, occlusion rate, distance from obstacle, and point score for each point in tabular form. The visualization solution generates a point distribution map overlaid with panoramic photos from drones, and marks the signal quality heat map and the line-of-sight paths between points; The field operation guide uses accessibility analysis based on drone aerial photography to generate the best arrival route for each point, avoiding dangerous areas.

[0014] Furthermore, the method also includes an iterative optimization step: The feature data obtained from the on-site survey and verification by UAVs are fed back to the control network agent as new samples to fine-tune the model and improve the initial selection accuracy of subsequent survey areas; The iterative mechanism includes updating positive and negative samples. Positive samples are control points that have passed the UAV verification, while negative samples are points that have been eliminated.

[0015] This application relates to the field of navigation and control technology, and in particular to a control network construction method based on dynamic optimization by an AI agent using a UAV. The method involves using a UAV equipped with a panoramic camera and a GNSS measurement module to collect control point data in real time. This data is then fed back to the AI ​​agent, which dynamically corrects the control point suitability factors and iteratively optimizes the control network points. The method includes a multi-factor dynamic evaluation and screening model for control points, and a mechanism for dynamically adjusting factor weights based on project objectives. For scenarios such as high mountains and dense forests, the method utilizes the UAV to identify temporary obstructions or terrain obstacles, and the AI ​​agent automatically adds candidate points to areas with accuracy gaps. This invention significantly improves the accuracy and reliability of the control network, substantially increases operational efficiency, shortens the control network design cycle, enhances adaptability to complex environments, and, through multi-source data fusion, can accurately address interference such as occlusion and multipath effects. Attached Figure Description

[0016] Figure 1 A flowchart illustrating the control network construction method based on dynamic optimization of UAV reconnaissance intelligent agents, as claimed in the embodiments of this application; Figure 2 A second flowchart of a control network construction method based on dynamic optimization of UAV reconnaissance intelligent agent, as claimed in an embodiment of this application; Figure 3The third flowchart is a control network construction method based on dynamic optimization of UAV reconnaissance intelligent agent, which is claimed in the embodiments of this application. Detailed Implementation

[0017] 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 a part of the embodiments of this application, and not all of the embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0018] The terms "first," "second," and "third" in this application are for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined as "first," "second," or "third" may explicitly or implicitly include at least one of that feature. In the description of this application, "multiple" means at least two, such as two, three, etc., unless otherwise explicitly specified. All directional indications (such as up, down, left, right, front, back, etc.) in the embodiments of this application are only used to explain the relative positional relationships and movements between components in a specific orientation (as shown in the figures). If the specific orientation changes, the directional indications also change accordingly. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or devices.

[0019] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a mutually exclusive, independent, or alternative embodiment. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0020] According to a first embodiment of the present invention, the present invention claims protection for a control network construction method based on dynamic optimization of UAV reconnaissance agents, referring to... Figure 1 This includes the following steps: S1: Construct a control network intelligent agent. The control network intelligent agent adopts a four-layer architecture of data layer, feature layer, model layer and application layer. It performs multi-source data fusion and analysis, candidate point adaptability scoring and preliminary selection result optimization on candidate control points in the survey area, and outputs a preliminary selection list of control points. S2: Based on the preliminary control point list, plan the UAV reconnaissance route, and use the UAV equipped with a panoramic high-definition camera, GNSS high-precision receiver and laser rangefinder to conduct on-site reconnaissance of each preliminary control point, and collect GNSS signal measurement data and environmental data. S3: Compare the measured data collected by the UAV with the candidate point information initially generated by the intelligent agent, and correct the satellite signal obstruction rate, PDOP value, terrain stability, spatial distribution rationality and interference signal source optimization factor; S4: Input the corrected optimization factors into the control network agent to perform dynamic optimization of the control network, including secondary evaluation to remove unqualified points, adding new points to optimize the distribution of the control network, and optimizing the connection relationship between points through the minimum spanning tree algorithm. S5: Output the optimal control network results, including a list of control points, a visualization scheme, and a field operation guide.

[0021] In this embodiment, the data collected by the UAV is compared with the candidate point information initially generated by the intelligent agent, and five core optimization factors are corrected, as shown in Table 1: Table 1 Optimization Factor Table

[0022] Furthermore, step S1 also includes: The data layer is used for inputting and preprocessing multi-source data, including terrain data, redline data, occlusion data, satellite signal prediction data, and auxiliary data, and performs coordinate system 1, ground filtering, and rasterization processing. The feature layer is used to extract key features from the preprocessed data, including redline avoidance, terrain stability, satellite signal quality, spatial distribution, road accessibility, and interference. The model layer employs a basic model combining a lightweight convolutional neural network and a gradient boosting tree, and integrates a rule engine for hard constraint verification. The application layer is used to output a list of coordinates of the initial control points and a visual map, and provides an iterative mechanism to fine-tune the model.

[0023] Furthermore, referring to Figure 2 The preprocessing steps of the data layer include: S121: Convert all input data to the 2000 National Geodetic Coordinate System to ensure spatial consistency of data; S122: Perform ground filtering on the LiDAR point cloud to separate ground points from vegetation or building points, and calculate the sky openness of each candidate point based on the filtered point cloud, wherein the sky openness is defined as the unobstructed zenith angle range. S123: Rasterize the satellite signal prediction data to generate a signal quality heatmap in grid cells, where each grid cell contains PDOP value, number of satellites and obstruction rate attributes. The multi-source data includes: High-resolution DEM data for slope analysis; The red line data, which consists of ecological protection red lines, permanent basic farmland, historical and cultural relic protection lines, and geological disaster risk lines, should be input in vector format. DSM data or LiDAR point clouds are used to extract tree and building height occlusion factors. Satellite visibility simulation data based on the latitude, longitude and topography of the survey area, including PDOP values ​​and satellite obstruction rates at different times; Auxiliary data consisting of traffic road data, geological disaster risk lines, high-voltage power transmission lines, and reservoir water surface distribution.

[0024] In this embodiment, the data layer input data and preprocessing include the following data: (1) Topographic data: High-resolution DEM data (slope analysis); (2) Red line data: data on red lines such as ecological protection red lines, permanent basic farmland, and historical and cultural relics protection lines (avoiding red line areas); (3) Occlusion data: DSM data or LiDAR point cloud (extracting occlusion factors such as tree and building height), images (identifying dense vegetation areas); (4) Satellite signal prediction data: Satellite visibility simulation based on the latitude, longitude and terrain of the survey area (such as PDOP value and satellite obstruction rate at different times). (5) Auxiliary data: traffic road data (accessibility of locations), geological disaster risk lines (avoiding landslide / debris flow areas), high-voltage power transmission lines, reservoir water surface distribution, etc. (affecting signals and multipath effects).

[0025] Preprocessing steps: Unify all data to the 2000 National Geodetic Coordinate System; Ground filtering is applied to the LiDAR point cloud to separate the ground from vegetation / buildings, and the "sky openness" (i.e., the unobstructed zenith angle range) of each candidate point is calculated. The satellite signal data is rasterized to generate a "signal quality heatmap". Furthermore, referring to Figure 3 The key feature extraction of the feature layer includes: S131: Extract red line avoidance features, and determine whether the coordinates of candidate points fall within the range of ecological protection red lines, permanent basic farmland, historical and cultural relics protection lines or geological disaster risk lines by superimposing the vector data of the survey area. If they are within the range, they are marked as red line violations. S132: Extract terrain stability features and calculate the slope and aspect of each candidate point based on DEM data, where the slope is in degrees and the aspect is in azimuth. S133: Extract satellite signal quality features. Obtain the number of satellites, PDOP value and occlusion rate corresponding to candidate points from the rasterized signal quality heat map. The occlusion rate is calculated by sky openness. S134: Extract spatial distribution features, calculate the distance between control points and the regional point density based on the coordinates of candidate points, where the point density is expressed as the number of points per square kilometer; S135: Extract road accessibility features, calculate the straight-line distance from candidate points to the nearest road based on traffic road vector data, and assign a difficulty score based on terrain complexity; S136: Extract interference features and calculate the straight-line distance between candidate points and high-voltage transmission lines and reservoir surfaces. If the distance is less than the threshold, it is marked as interference risk.

[0026] In this embodiment, the feature layer is shown in Table 2; Table 2 Key Features Table Feature categories Specific features Physical meaning Red line avoidance Outside / inside the ecological protection red line, permanent basic farmland, historical and cultural relic protection line, and geological disaster risk line Avoid encroaching on the red line Terrain stability Slope (°), Aspect (°) Assess the foundation stability of the site, avoiding steep slopes / loose areas. Satellite signal quality Number of satellites, PDOP value, and obstruction rate (%) A GNSS observation quality rating of ≤30% is considered excellent. Spatial distribution <![CDATA[Distance between control points (km), regional point density (points / km 2 )]]> Ensure the mesh is uniform, avoiding it being too dense or too sparse. Road accessibility Distance from road (m), difficulty of passage rating To reduce field setup costs, locations ≤500m from roads are recommended. Interference Distance from high-voltage power line (m), distance from reservoir surface (m) Keep away from sources of signal interference and avoid multipath effects. Furthermore, the specific implementation of the model layer includes: A lightweight convolutional neural network is used to process rasterized signal quality heatmaps and terrain data to extract spatial features; Gradient boosting trees are used to process structured features, including slope, distance, and density, and output the point scores of candidate points. The rule engine performs hard constraint verification on the points output by the model, forcibly removing points with a PDOP value greater than 6, a number of satellites less than 4, or located within the red line range, and conditionally removing points with a spacing of less than 500 meters and redundant signal quality.

[0027] In this embodiment, the model layer adopts a hybrid model of "deep learning + rule constraints" to balance prediction accuracy and engineering rationality. A lightweight convolutional neural network (CNN) + gradient boosting tree (XGBoost) is used. CNN is used to process the rasterized signal quality heatmap and terrain data to extract spatial features. XGBoost is used to process structured features (such as slope and distance) and output the "location score" of individual candidate points.

[0028] Positive samples in the training sample: control points that have been verified as qualified in previous similar projects (with actual observation accuracy data). Negative samples: Candidate points that are removed (such as points where observations failed due to occlusion, or points with unstable terrain); Tags: Marked based on drone survey findings.

[0029] The rules engine supplements the model output with hard constraint verification based on UAV survey data, eliminating points that are "misjudged by the model": Forced exclusion: Points with PDOP value > 6, number of satellites < 4, or located within the red line area; Elimination criteria: If the distance between two points is less than 500m and both are high resolution, retain the point with better signal quality (to avoid redundancy).

[0030] Furthermore, the specific steps of the on-site reconnaissance by UAV in step S2 include: Step S21: Plan the UAV flight path based on the preliminary control point list, ensuring that the flight path covers all candidate points with a score of 70 or higher; Step S22: The UAV flies to each candidate point in sequence, hovers at a height of 15-2 meters, simulates the GNSS receiver environment, hovers for 10-15 minutes to collect GNSS data, including 10 sets of single-point measurements, smoothing each set 10 times, and recording the number of satellites, PDOP value, carrier-to-noise ratio and multipath error; Step S23: Use a 360-degree panoramic high-definition camera to take panoramic photos of each point. The gimbal rotates one frame every 30 degrees, and the resolution is no less than 12 million pixels. This is used to verify vegetation obstruction, terrain changes and sources of interference. Step S24: Obtain the three-dimensional coordinates of the UAV through the GNSS receiver, and combine them with the altitude measured by the laser rangefinder to geometrically calculate the three-dimensional coordinates of the preset control point; Step S25: Transmit the panoramic photos, GNSS data, and coordinate information to the computer processing software in real time.

[0031] Furthermore, step S3 also includes: For satellite signal obstruction rate, compare the actual obstruction rate of the UAV with the predicted value of the agent. If the actual value is more than 15% higher than the predicted value, the score of this factor will be directly reduced. For PDOP values, if the measured PDOP value of the UAV is greater than 6 and the predicted value is less than or equal to 6, it is marked as a signal risk point and the PDOP factor is corrected. For terrain stability, the slope value calculated by DEM is corrected based on the detailed slope and foundation condition captured by drone. If loose foundation or hidden steep slope is found, the score is adjusted. To ensure the rationality of spatial distribution, the theoretical point spacing was adjusted based on the actual terrain obstacles captured by drone aerial photography. For sources of interference, the actual distance to high-voltage lines or reservoirs is determined based on high-definition images from drones. If the images show an unpredictable source of interference, the distance value is adjusted.

[0032] In this embodiment, the scoring algorithm is as follows: In line with the requirements of "red line prohibition" and "core factor score control", the following is a multi-factor evaluation system design for the selection of control point sites, which includes veto items (rigid exclusion), core factor scoring standards, penalty conditions and comprehensive scoring formula, taking into account both compliance and practicality.

[0033] For areas that are legally prohibited or absolutely unusable, there is no need to proceed to the scoring stage; they are directly judged as "unqualified." Within the ecological protection red line: located within the ecological protection red line area designated by governments at all levels (in accordance with the "Ecological Protection Red Line Management Measures"); Permanent basic farmland: Areas included in the permanent basic farmland database (based on national land spatial planning); No-construction zones within urban development boundaries: such as ecological corridors and core areas of water source protection zones as clearly defined in the plan; Other legally prohibited areas include military restricted areas and core areas for cultural relic protection.

[0034] Judgment method: By overlaying the vector data of the survey area (red line shp file), if the coordinates of the candidate point fall within the above area, it is directly excluded.

[0035] To address the defect of "approaching disqualification but not meeting the veto criteria," additional deductions are implemented to prevent "gray area" situations. Critical satellite signal obstruction: If the obstruction rate is 40%-50% (approaching severe obstruction), an additional 5 points will be deducted; Road accessibility threshold: If the distance to the road is 800-1000m (difficult to pass), an additional 3 points will be deducted; Terrain risk hazards: If the slope is 30°-35° (close to a steep slope) and located on the edge of a geological disaster-prone area, an additional 4 points will be deducted.

[0036] Deduction rules: Triggering one item will deduct the corresponding points, which can be accumulated (until the total score reaches 0 points).

[0037] Furthermore, step S4 also includes: Step S41: Secondary evaluation and elimination. The candidate point fit is recalculated using the corrected factors. Points with a measured PDOP value greater than 6 and lasting for more than 5 minutes, actual occlusion rate greater than 50%, or less than 50 meters from the high-voltage line or less than 200 meters from the reservoir surface are eliminated. Step S42: Gap filling. For the accuracy gap area that appears after elimination, add new candidate points within a 1-2 kilometer radius, prioritizing areas that are verified by UAV as having low occlusion and a PDOP value of less than or equal to 6. Step S43: Distribution optimization. The minimum spanning tree algorithm is used to optimize the connection relationship between the retained points to ensure that there are no obstacles confirmed by the UAV between adjacent points and that the average spacing meets the accuracy requirements of the control network. Step S44: Multi-objective balance verification to ensure that the final control network has a score of 70 or higher, the maximum PDOP of the entire network is less than or equal to 6, the total number of points decreases by no more than 15%, and at least 30% of the points can be used as benchmark points.

[0038] In this embodiment, the comprehensive scoring formula and result determination are... (1) Calculation formula:

[0039] Where: m is the number of penalty conditions triggered; If a veto is triggered, the final result will be directly judged as "unqualified" (no score will be calculated).

[0040] Result Judgment Criteria Qualified: Point score ≥ 70 points (no veto items, and no serious defects in core factors); Priority candidates: Site score ≥ 90 points (number of satellites ≥ 10 and PDOP ≤ 6, satellite obstruction rate ≤ 10%, road distance ≤ 300m, slope ≤ 15°); Unqualified: The point score is less than 70 points or any veto item is triggered.

[0041] The application layer includes result output and iteration: Output format: List of coordinates of preliminary control points (with point scores and key feature values); Visual map (marking the location of candidate points and the reasons for rejection).

[0042] Iterative mechanism: After conducting on-site reconnaissance and verification of the initial selected points using drones, the feature data is fed back to the agent, and the model is fine-tuned using new samples to improve the initial selection accuracy of subsequent survey areas. Furthermore, the output of the optimal control network in step S5 includes: The control point list lists the coordinates, measured PDOP value, occlusion rate, distance from obstacle, and point score for each point in tabular form. The visualization solution generates a point distribution map overlaid with panoramic photos from drones, and marks the signal quality heat map and the line-of-sight paths between points; The field operation guide uses accessibility analysis based on drone aerial photography to generate the best arrival route for each point, avoiding dangerous areas.

[0043] Furthermore, the method also includes an iterative optimization step: The feature data obtained from the on-site survey and verification by UAVs are fed back to the control network agent as new samples to fine-tune the model and improve the initial selection accuracy of subsequent survey areas; The iterative mechanism includes updating positive and negative samples. Positive samples are control points that have passed the UAV verification, while negative samples are points that have been eliminated.

[0044] In this embodiment, the drone survey includes using drones to survey the on-site conditions of control points and planning drone survey routes based on a pre-selected control point database with scores of ≥70.

[0045] Acquire measured GNSS signal data. The UAV sequentially flies to each initially selected control point and hovers over the candidate points for 10-15 minutes (the UAV hovering height is the design control point observation tower height, typically 1.5-2 meters). Simulate multi-satellite and single BeiDou system environments for the GNSS receiver, and collect GNSS data (10 sets of single-point measurements, smoothed 10 times per set). Parameters include the number of satellites received, PDOP value, etc. Specifically, this includes: number of satellites, effective satellites per minute (≥5); PDOP value, position accuracy attenuation factor (≤3, lower is better); signal strength, carrier-to-noise ratio of each satellite; multipath error, MP1 / MP2 values ​​recorded by the GNSS module.

[0046] Acquire panoramic environmental data. Use a drone's rotatable angle camera to capture 360-degree panoramic high-definition photos (resolution ≥ 12 megapixels) at each point. Rotate the gimbal to capture panoramic photos (one frame per 30°). Re-verify the accuracy of the initial selection conditions, mainly vegetation obstruction, geological disaster conditions, traffic accessibility conditions, and signal analysis conditions. Identify actual obstructions (such as temporary buildings and dense foliage not reflected in previous LiDAR data) and terrain changes (such as recent landslides and new obstacles). Calculate the actual "sky open angle" through perspective analysis of the panoramic photos (correcting previous satellite obstruction rate predictions). Verify terrain slope and geological stability (such as the presence of loose rocks and water accumulation). Record the actual distance to high-voltage lines and reservoirs (by comparing RTK positioning coordinates with vector data).

[0047] The 3D coordinates of the UAV are obtained using a GNSS receiver, and the height of the UAV from the preset control point is obtained using a laser rangefinder. The 3D coordinates of the preset control point are then calculated using geometric relationships.

[0048] The panoramic photos, GNSS data, and 3D coordinates of preset control points are fed back to the computer processing software. The AI ​​agent then dynamically optimizes and adjusts the control network based on the survey data returned by the drone. The AI ​​agent's dynamic optimization control network inputs the corrected factor data into the agent, initiating a "secondary evaluation-elimination-supplementation" optimization process: A secondary evaluation and elimination process is employed. The agent recalculates the fit of candidate points using updated factor scores (such as satellite signal and PDOP values) and eliminates the following points: Actual measured PDOP > 6 and lasting for more than 5 minutes (poor signal stability); The actual occlusion rate is greater than 50% (panoramic photos confirm the existence of dense occlusion). <50 meters from high-voltage lines or <200 meters from the surface of a reservoir (violation of multipath constraints).

[0049] To address gap filling and distribution optimization, for the "precision gap area" that appears after elimination (such as the area originally covered by 3 control points but now only 1 remains), the agent adds new candidate points within a 1-2 kilometer radius based on the terrain data measured by UAVs (prioritizing areas verified by UAVs as "low occlusion, PDOP≤6"). The connection relationship of the retained points is optimized by using the "minimum spanning tree" algorithm to ensure that there are no measured obstacles between adjacent points (such as the line-of-sight path confirmed by the UAV) and that the average spacing meets the accuracy requirements of the control network.

[0050] The control network output by the intelligent agent must satisfy the following: All points were verified by drone testing (no disqualifications, score ≥ 70 points). The maximum PDOP value across the entire network is ≤6 (to ensure overall accuracy). The total number of points is reduced by ≤15% compared to the initial plan (to control costs); At least 30% of the points can be used as reference points.

[0051] The final generated optimal control network contains three types of outcomes: Control point list: including key parameters such as coordinates, measured PDOP value, occlusion rate, and distance from obstacle; Visualization solution: Overlay a point distribution map from the drone panoramic photo, and label each point with a "signal quality heatmap" and line-of-sight path; Fieldwork Guidelines: Based on accessibility analysis using drone aerial photography, provide the optimal route to each point (avoiding hazardous areas identified by the drone).

[0052] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces, or indirect coupling or communication connection between apparatuses or units, and may be electrical, mechanical, or other forms.

[0053] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated units described above can be implemented in hardware or as software functional units. The above are merely embodiments of this application and do not limit the patent scope of this application. Any equivalent structural or procedural transformations made based on the description and drawings of this application, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

[0054] The specific embodiments of the invention have been described in detail above, but they are only examples, and this application is not limited to the specific embodiments described above. For those skilled in the art, any equivalent modifications or substitutions to the invention are also within the scope of this application. Therefore, all equivalent changes, modifications, and improvements made without departing from the spirit and principles of this application should be covered within the scope of this application.

Claims

1. A method for constructing a control network based on dynamic optimization of an unmanned aerial vehicle (UAV) reconnaissance agent, characterized in that, Includes the following steps: S1: Construct a control network intelligent agent. The control network intelligent agent adopts a four-layer architecture of data layer, feature layer, model layer and application layer. It performs multi-source data fusion and analysis, candidate point adaptability scoring and preliminary selection result optimization on candidate control points in the survey area, and outputs a preliminary selection list of control points. S2: Based on the preliminary control point list, plan the UAV reconnaissance route, and use the UAV equipped with a panoramic high-definition camera, GNSS high-precision receiver and laser rangefinder to conduct on-site reconnaissance of each preliminary control point, and collect GNSS signal measurement data and environmental data. S3: Compare the measured data collected by the UAV with the candidate point information initially generated by the intelligent agent, and correct the satellite signal obstruction rate, PDOP value, terrain stability, spatial distribution rationality and interference signal source optimization factor; S4: Input the corrected optimization factors into the control network agent to perform dynamic optimization of the control network, including secondary evaluation to remove unqualified points, adding new points to optimize the distribution of the control network, and optimizing the connection relationship between points through the minimum spanning tree algorithm. S5: Output the optimal control network results, including a list of control points, a visualization scheme, and a field operation guide; Step S1 also includes: The data layer is used for inputting and preprocessing multi-source data, including terrain data, redline data, occlusion data, satellite signal prediction data, and auxiliary data, and performs coordinate system 1, ground filtering, and rasterization processing. The feature layer is used to extract key features from the preprocessed data, including redline avoidance, terrain stability, satellite signal quality, spatial distribution, road accessibility, and interference. The model layer employs a basic model combining a lightweight convolutional neural network and a gradient boosting tree, and integrates a rule engine for hard constraint verification. The application layer is used to output a list of coordinates of the initial control points and a visual map, and provides an iterative mechanism to fine-tune the model. Step S3 also includes: For satellite signal obstruction rate, compare the actual obstruction rate of the UAV with the predicted value of the agent. If the actual value is more than 15% higher than the predicted value, the score of this factor will be directly reduced. For PDOP values, if the measured PDOP value of the UAV is greater than 6 and the predicted value is less than or equal to 6, it is marked as a signal risk point and the PDOP factor is corrected. For terrain stability, the slope value calculated by DEM is corrected based on the detailed slope and foundation condition captured by drone. If loose foundation or hidden steep slope is found, the score is adjusted. To ensure the rationality of spatial distribution, the theoretical point spacing was adjusted based on the actual terrain obstacles captured by drone aerial photography. For sources of interference signals, the actual distance to high-voltage lines or reservoirs is determined based on high-definition images from drones. If the images show unpredicted sources of interference, the distance value is adjusted. Step S4 also includes: Step S41: Secondary evaluation and elimination. The candidate point fit is recalculated using the corrected factors. Points with a measured PDOP value greater than 6 and lasting for more than 5 minutes, actual occlusion rate greater than 50%, or less than 50 meters from the high-voltage line or less than 200 meters from the reservoir surface are eliminated. Step S42: Gap filling. For the accuracy gap area that appears after elimination, add new candidate points within a 1-2 kilometer radius, prioritizing areas that are verified by UAV as having low occlusion and a PDOP value of less than or equal to 6. Step S43: Distribution optimization. The minimum spanning tree algorithm is used to optimize the connection relationship between the retained points to ensure that there are no obstacles confirmed by the UAV between adjacent points and that the average spacing meets the accuracy requirements of the control network. Step S44: Multi-objective balance verification to ensure that the final control network has a score of 70 or higher, the maximum PDOP of the entire network is less than or equal to 6, the total number of points decreases by no more than 15%, and at least 30% of the points can be used as benchmark points.

2. The control network construction method based on dynamic optimization of UAV reconnaissance agent according to claim 1, characterized in that, The preprocessing steps of the data layer include: S121: Convert all input data to the 2000 National Geodetic Coordinate System to ensure spatial consistency of data; S122: Perform ground filtering on the LiDAR point cloud to separate ground points from vegetation or building points, and calculate the sky openness of each candidate point based on the filtered point cloud, wherein the sky openness is defined as the unobstructed zenith angle range. S123: Rasterize the satellite signal prediction data to generate a signal quality heatmap in grid cells, where each grid cell contains PDOP value, number of satellites and obstruction rate attributes. The multi-source data includes: High-resolution DEM data for slope analysis; The red line data, which consists of ecological protection red lines, permanent basic farmland, historical and cultural relic protection lines, and geological disaster risk lines, should be input in vector format. DSM data or LiDAR point clouds are used to extract tree and building height occlusion factors. Satellite visibility simulation data based on the latitude, longitude and topography of the survey area, including PDOP values ​​and satellite obstruction rates at different times; Auxiliary data consisting of traffic road data, geological disaster risk lines, high-voltage power transmission lines, and reservoir water surface distribution.

3. The control network construction method based on dynamic optimization of UAV reconnaissance intelligent agent according to claim 1, characterized in that, The key feature extraction of the feature layer includes: S131: Extract red line avoidance features, and determine whether the coordinates of candidate points fall within the range of ecological protection red lines, permanent basic farmland, historical and cultural relics protection lines or geological disaster risk lines by superimposing the vector data of the survey area. If they are within the range, they are marked as red line violations. S132: Extract terrain stability features and calculate the slope and aspect of each candidate point based on DEM data, where the slope is in degrees and the aspect is in azimuth. S133: Extract satellite signal quality features. Obtain the number of satellites, PDOP value and occlusion rate corresponding to candidate points from the rasterized signal quality heat map. The occlusion rate is calculated by sky openness. S134: Extract spatial distribution features, calculate the distance between control points and the regional point density based on the coordinates of candidate points, where the point density is expressed as the number of points per square kilometer; S135: Extract road accessibility features, calculate the straight-line distance from candidate points to the nearest road based on traffic road vector data, and assign a difficulty score based on terrain complexity; S136: Extract interference features and calculate the straight-line distance between candidate points and high-voltage transmission lines and reservoir surfaces. If the distance is less than the threshold, it is marked as interference risk.

4. The control network construction method based on dynamic optimization of UAV reconnaissance agent according to claim 1, characterized in that, The specific implementation of the model layer includes: A lightweight convolutional neural network is used to process rasterized signal quality heatmaps and terrain data to extract spatial features; Gradient boosting trees are used to process structured features, including slope, distance, and density, and output the point scores of candidate points. The rule engine performs hard constraint verification on the points output by the model, forcibly removing points with a PDOP value greater than 6, a number of satellites less than 4, or located within the red line range, and conditionally removing points with a spacing of less than 500 meters and redundant signal quality.

5. The control network construction method based on dynamic optimization of UAV reconnaissance intelligent agent according to claim 1, characterized in that, The specific steps of the on-site reconnaissance by UAV in step S2 include: Step S21: Plan the UAV flight path based on the preliminary control point list, ensuring that the flight path covers all candidate points with a score of 70 or higher; Step S22: The UAV flies to each candidate point in sequence, hovers at a height of 1.5-2 meters, simulates the GNSS receiver environment, hovers for 10-15 minutes to collect GNSS data, including 10 sets of single-point measurements, smoothing each set 10 times, and recording the number of satellites, PDOP value, carrier-to-noise ratio and multipath error; Step S23: Use a 360-degree panoramic high-definition camera to take panoramic photos of each point. The gimbal rotates one frame every 30 degrees, and the resolution is no less than 12 million pixels. This is used to verify vegetation obstruction, terrain changes and sources of interference. Step S24: Obtain the three-dimensional coordinates of the UAV through the GNSS receiver, and combine them with the altitude measured by the laser rangefinder to geometrically calculate the three-dimensional coordinates of the preset control point; Step S25: Transmit the panoramic photos, GNSS data, and coordinate information to the computer processing software in real time.

6. The control network construction method based on dynamic optimization of UAV reconnaissance intelligent agent according to claim 1, characterized in that, The output of the optimal control network in step S5 includes: The control point list lists the coordinates, measured PDOP value, occlusion rate, distance from obstacle, and point score for each point in tabular form. The visualization solution generates a point distribution map overlaid with panoramic photos from drones, and marks the signal quality heat map and the line-of-sight paths between points; The field operation guide uses accessibility analysis based on drone aerial photography to generate the best arrival route for each point, avoiding dangerous areas.

7. The control network construction method based on dynamic optimization of UAV reconnaissance intelligent agent according to claim 1, characterized in that, It also includes iterative optimization steps: The feature data obtained from the on-site survey and verification by UAVs are fed back to the control network agent as new samples to fine-tune the model and improve the initial selection accuracy of subsequent survey areas; The iterative mechanism includes updating positive and negative samples. Positive samples are control points that have passed the UAV verification, while negative samples are points that have been eliminated.

Images