Intelligent optimization decision support system for flight path based on low-altitude digital twin model

Abstract

The present application relates to the field of low-altitude flight control and digital twin technology, and in particular to a flight path intelligent optimization decision support system based on a low-altitude digital twin model, which comprises a multi-source data perception fusion module, a low-altitude digital twin modeling module, an intelligent path optimization decision module, and a virtual-real linkage control feedback module.A multi-scale dynamic digital twin model is constructed to realize real-time high-precision mapping of static terrain and dynamic environment.A three-stage path optimization algorithm combining imitation learning and reinforcement learning is proposed to complete closed-loop training in combination with a digital twin simulation environment.A distributed federated reinforcement learning collaboration mechanism is designed to ensure the safety and privacy of multi-aircraft path collaborative optimization.A space-time evolution type risk assessment model is established to realize early warning of dynamic risks and adaptive adjustment of paths.The present application can significantly improve the safety, economy and real-time performance of low-altitude flight paths.

Description

Technical Field

[0001] This invention relates to the field of low-altitude flight control and digital twin technology, specifically to a flight path intelligent optimization decision support system based on a low-altitude digital twin model. Background Technology

[0002] With the rapid development of the low-altitude economy, low-altitude aircraft such as drones are increasingly being used in logistics delivery, power line inspection, emergency rescue, and urban monitoring. The planning and optimization of low-altitude flight paths are core technologies for ensuring flight safety and improving mission efficiency. Existing low-altitude flight path planning technologies mainly suffer from the following shortcomings: 1. Lagging and insufficient accuracy in environmental perception: Existing technologies mostly rely on single sensors or environmental data from fixed time periods, making it difficult to capture changes in the dynamic low-altitude environment (such as sudden weather changes, temporary obstacles, and abrupt changes in the trajectories of other aircraft) in real time. This leads to deviations between path planning and the actual environment, increasing flight risks. For example, some solutions only integrate static terrain and preset meteorological data for path planning, which cannot cope with sudden weather events such as real-time wind shear and short-term heavy rainfall.

[0003] 2. Poor adaptability of path optimization algorithms: Existing path optimization algorithms mostly employ traditional heuristic algorithms (such as genetic algorithms and particle swarm optimization algorithms) or single reinforcement learning algorithms, which suffer from problems such as cold start in training, causal confusion, and large differences between open-loop training and closed-loop deployment. Traditional heuristic algorithms have slow convergence speed in complex dynamic environments, making it difficult to meet real-time optimization requirements; single reinforcement learning algorithms are prone to local optima and have high training costs and poor security in real-world environments.

[0004] 3. Inefficient and privacy-insensitive multi-aircraft collaborative decision-making: In multi-aircraft collaborative flight scenarios, existing technologies mostly adopt centralized data processing and decision-making models, which suffer from problems such as data transmission delays and central node computing power bottlenecks, and are prone to leaking the mission privacy and location information of each aircraft. Some distributed collaborative solutions lack effective model aggregation mechanisms, resulting in low collaborative decision-making accuracy and a tendency for path conflicts.

[0005] 4. Insufficient integration of virtual and real worlds: Existing technologies lack a sound mechanism for linking the physical world and virtual models. It is impossible to verify the safety and feasibility of the path in advance through virtual simulation, and it is also difficult to dynamically correct and optimize the model based on real flight data, resulting in insufficient reliability and robustness of path planning.

[0006] Digital twin technology, as a key technology for achieving accurate mapping and real-time interaction between the physical and virtual worlds, provides a new approach to solving the aforementioned problems in low-altitude flight path planning. However, existing low-altitude flight-related technologies combined with digital twins still suffer from problems such as loose coupling between the digital twin model and the path optimization algorithm, lagging model updates, and significant differences between the simulation environment and the real environment. A complete closed-loop system of "perception-modeling-optimization-feedback-iteration" has not yet been formed.

[0007] Therefore, there is an urgent need to design a decision support system for intelligent optimization of low-altitude flight paths that integrates digital twins and advanced intelligent algorithms to improve the safety, efficiency and reliability of low-altitude flight. Summary of the Invention

[0008] To address the problems of lagging environmental perception, poor adaptability of path optimization, inefficient multi-aircraft collaborative decision-making, and insufficient virtual-real linkage in existing low-altitude flight path planning technologies, this invention provides a flight path intelligent optimization decision support system based on a low-altitude digital twin model. Through a combination of innovative technologies such as multi-scale dynamic digital twin modeling, three-stage reinforcement learning optimization, distributed federated collaborative decision-making, and full-process virtual-real linkage feedback, it achieves real-time, accurate, and safe optimization of low-altitude flight paths, thereby improving the overall efficiency of low-altitude flight missions.

[0009] The technical solution adopted by this invention to solve its technical problem is: a flight path intelligent optimization decision support system based on a low-altitude digital twin model, comprising: a multi-source data perception and fusion module, used to collect static and dynamic data of the low-altitude flight area, and complete data preprocessing and spatiotemporal synchronization through an adaptive weighted fusion algorithm; a low-altitude digital twin modeling module, which constructs a digital twin of the low-altitude airspace based on multi-scale dynamic modeling technology, realizing real-time mapping and dynamic updating of the physical airspace and the virtual model; an intelligent path optimization decision module, which adopts a three-stage training architecture reinforcement learning algorithm, combined with a digital twin simulation environment to complete the training and inference of the path optimization model, and outputs the initial optimized path; and a multi-aircraft collaborative decision module, based on a credibility-weighted distributed federated reinforcement learning mechanism. The system achieves collaborative optimization and conflict avoidance of multi-vehicle paths; a virtual-real linkage control feedback module collects real-time flight status data of the aircraft and feeds it back to the digital twin model for deviation correction, driving the path optimization model to dynamically adjust path parameters; the low-altitude digital twin modeling module uses BeiDou grid spatial coding technology to achieve airspace grid management, integrates 3DGS technology to improve model rendering accuracy, and adopts an accuracy adaptive adjustment strategy to dynamically match modeling accuracy according to flight mission requirements; the three-stage training architecture includes a perception pre-training stage, an imitation learning initialization stage, and a reinforcement learning fine-tuning stage, guiding the model to learn the causal relationships of the physical world through a safety-oriented reward function, which comprehensively considers collision risk, trajectory deviation, energy consumption cost, and mission completion efficiency.

[0010] Specifically, the static data collected by the multi-source data perception and fusion module includes terrain elevation data, building outline data, take-off and landing point distribution data, and airspace control boundary data; the dynamic data includes real-time meteorological data, aircraft status data, dynamic obstacle data, and communication link status data; the adaptive weighted fusion algorithm dynamically adjusts the weights of each sensor data based on the entropy weight assignment method, and combines the collaborative Kalman filter algorithm to eliminate data noise and redundancy.

[0011] Specifically, the dynamic update mechanism of the low-altitude digital twin modeling module includes: processing multi-source sensing data in real time based on edge computing nodes and updating the environmental parameters of the digital twin model according to a preset time step; when a sudden dynamic obstacle or meteorological change is detected, an emergency update process is triggered to prioritize updating the model data of key areas that affect flight safety.

[0012] Specifically, the safety-oriented reward function is expressed as: R1=α·R2+β·R3-γ·R4-δ·R5; where R2 is the safety reward, which is a positive reward when there is no collision risk and a negative penalty when a collision risk is detected; R3 is the task efficiency reward, which is negatively correlated with the task completion time; R4 is the energy cost penalty, which is positively correlated with flight energy consumption; R5 is the trajectory deviation penalty, which is positively correlated with the deviation distance of the preset reference trajectory; α, β, γ, and δ are weight adjustment parameters that satisfy α+β+γ+δ=1.

[0013] Specifically, the distributed federated reinforcement learning mechanism includes: each aircraft acts as a local training node, completing local model training based on its own perception data and digital twin simulation data; the credibility of each local node's model is calculated through a credibility evaluation module, and the credibility is determined based on the quality of training data and the accuracy of model prediction; the parameters of each local model are aggregated according to the credibility weight to generate a global optimization model, thereby achieving multi-aircraft collaborative decision-making while protecting data privacy.

[0014] Specifically, the virtual-real linkage control feedback module includes a deviation detection unit and a path adjustment unit; the deviation detection unit calculates the deviation value between the actual flight trajectory of the aircraft and the predicted trajectory of the digital twin model, and triggers the path adjustment unit when the deviation value exceeds a preset threshold; the path adjustment unit, based on an incremental learning algorithm, integrates real-time deviation data and environmental change data to perform local fine-tuning of the optimized path.

[0015] Specifically, it also includes a spatiotemporal evolution-based risk assessment module. This module constructs a risk prediction model based on a Bayesian network, integrates historical flight data, real-time environmental data, and aircraft status data, predicts the risk evolution trend of low-altitude airspace within a preset time period, and outputs a risk level map to provide risk-oriented basis for path optimization decisions.

[0016] Specifically, the risk level map adopts a grid-based division method, and the risk level of each grid unit is determined by a combination of static risk factors and dynamic risk factors; static risk factors include terrain complexity and building density; dynamic risk factors include the probability of sudden weather changes, the probability of dynamic obstacles appearing, and the probability of communication interruption.

[0017] Specifically, the intelligent path optimization decision module also includes a multi-objective optimization constraint unit, which includes minimum safe altitude constraint, maximum flight speed constraint, energy consumption threshold constraint, and mission time window constraint. The conflict between the various constraints is coordinated through a hierarchical constraint processing mechanism.

[0018] Specifically, the low-altitude digital twin modeling module also integrates a scene simulation unit, which can simulate different weather conditions, obstacle distributions, and aircraft malfunction scenarios, providing a diverse simulation environment for the training and verification of the path optimization model.

[0019] The beneficial effects of this invention are: 1. Significantly improved environmental perception accuracy and real-time performance: Through multi-source data fusion and adaptive weighting algorithms, full-dimensional and high-precision perception of the low-altitude environment is achieved; combined with edge computing and emergency update mechanisms, the real-time synchronization between the digital twin model and the physical world is ensured, solving the problem of lagging environmental perception in existing technologies.

[0020] 2. Enhanced Adaptability and Reliability of Path Optimization: The innovative three-stage training architecture combined with the digital twin simulation environment avoids the cold start and causal confusion problems of reinforcement learning, and improves the model's adaptability in complex dynamic environments; the safety-oriented reward function and multi-objective constraint processing mechanism ensure that path optimization takes into account safety, efficiency and economy, and reduces flight risks.

[0021] 3. High efficiency and privacy security in multi-vehicle collaborative decision-making: The distributed federated reinforcement learning mechanism solves the latency and computing power bottleneck problems of centralized collaborative decision-making, while protecting the mission privacy and location information of each aircraft; the credibility weighted aggregation ensures the accuracy of the global collaborative model and effectively avoids multi-vehicle path conflicts.

[0022] 4. Form a complete closed-loop optimization system: Through the virtual-real linkage control feedback module, the physical flight data is used to dynamically correct the virtual model and optimization algorithm, forming a complete closed loop of "perception-modeling-optimization-feedback-iteration", which continuously improves the reliability and robustness of path planning.

[0023] 5. Wide range of applications: It can be adapted to various low-altitude flight missions such as logistics distribution, power line inspection, emergency rescue, and urban monitoring. Through precision adaptive adjustment and scenario simulation functions, it can meet the personalized needs of different missions and has extremely high practical value and promotion prospects. Attached Figure Description

[0024] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0025] Figure 1 This is an overall architecture diagram of the intelligent flight path optimization decision support system based on a low-altitude digital twin model, as described in this invention. Figure 2 This is a flowchart illustrating the workflow of the multi-source data sensing and fusion module of the present invention. Figure 3 This is a schematic diagram of the multi-scale dynamic modeling of the low-altitude digital twin modeling module of the present invention; Figure 4 This is a diagram of the three-stage training architecture of the intelligent path optimization decision module of the present invention; Figure 5 This is a schematic diagram of the distributed federated reinforcement learning of the multi-aircraft collaborative decision-making module of the present invention; Figure 6 This is a flowchart of the virtual-real linkage control feedback module of the present invention. Detailed Implementation

[0026] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below in conjunction with specific embodiments.

[0027] like Figures 1-6 As shown, the intelligent flight path optimization decision support system based on a low-altitude digital twin model described in this invention includes a multi-source data perception and fusion module, a low-altitude digital twin modeling module, an intelligent path optimization decision module, a multi-aircraft collaborative decision module, a virtual-real linkage control feedback module, and a spatiotemporal evolution risk assessment module. These modules work collaboratively to form a complete closed-loop optimization system, specifically including the following modules: 1. Multi-source data sensing and fusion module This module is the foundation for the system to perceive the physical environment. It is responsible for collecting full-dimensional data of the low-altitude flight area and completing the data preprocessing and spatiotemporal synchronization.

[0028] (1) Data Acquisition: The acquired data is divided into static data and dynamic data. Static data includes terrain elevation data (accuracy better than 0.5m), building outline data (including height and structure type), take-off and landing point distribution data (including location, load-bearing capacity, take-off and landing restrictions) and airspace control boundary data (including no-fly zones, restricted flight zones, and flight altitude restrictions) obtained through satellite remote sensing and lidar scanning; dynamic data includes aircraft status data (position, speed, attitude, and battery level) collected through airborne sensors (IMU, GNSS, visual inertial navigation), real-time meteorological data (wind speed, wind direction, rainfall, visibility, and air pressure) collected through meteorological sensor networks, dynamic obstacle data (vehicles, pedestrians, other aircraft, and temporary construction areas) obtained through video surveillance and radar detection, and communication link status data (bandwidth, delay, and packet loss rate) collected through communication link monitoring units.

[0029] (2) Data preprocessing and fusion: An adaptive weighting algorithm based on entropy weight assignment is adopted to dynamically evaluate the reliability of each sensor data and adjust the data weight; the sensor data is denoised by wavelet transform algorithm to retain key information; the spatiotemporal synchronization and fusion of multi-source data is realized by using collaborative Kalman filtering algorithm to eliminate data redundancy and conflict, and output high-precision and high-reliability unified environmental data and aircraft status data.

[0030] 2. Low-altitude digital twin modeling module This module constructs a digital twin of the physical low-altitude airspace, enabling real-time and accurate mapping between the physical and virtual worlds, and providing a virtual platform for path optimization and simulation verification.

[0031] (1) Multi-scale dynamic modeling: Based on BeiDou grid spatial coding technology, the low-altitude flight area is divided into grid units of different scales (such as 1m×1m high-precision grid, 10m×10m conventional grid, and 50m×50m large-scale grid). 3DGS (3D Gaussian Splatting) technology is integrated to construct a high-fidelity 3D model, realizing refined rendering of static elements such as terrain, buildings, and take-off and landing points; multi-source dynamic data is integrated to realize real-time mapping of dynamic elements such as meteorological fields, dynamic obstacles, and aircraft trajectories. An accuracy adaptive adjustment strategy is adopted to dynamically match the modeling accuracy according to the flight mission type (such as 1m×1m grid for high-precision inspection missions and 10m×10m grid for large-scale logistics delivery missions) and the complexity of the flight area, balancing model accuracy and computational efficiency.

[0032] (2) Dynamic update mechanism: Real-time data processing units are deployed based on edge computing nodes to update the environmental parameters of the digital twin model according to a preset time step (1s for normal scenarios and 0.2s for complex dynamic scenarios); emergency update trigger conditions are set. When a sudden dynamic obstacle (such as a vehicle or pedestrian entering the airspace), a sudden change in weather (such as wind shear or short-term heavy rainfall) or a communication link interruption is detected, the emergency update process is triggered to prioritize updating the model data of key areas that affect flight safety, so as to ensure the consistency between the digital twin model and the physical world.

[0033] (3) Scene simulation unit: It integrates diverse scene simulation functions, which can simulate different meteorological conditions (heavy rain, strong wind, low visibility), obstacle distribution (dense crowds, complex building clusters) and aircraft failure scenarios (insufficient power, sensor failure, communication interruption), providing a diverse virtual simulation environment for model training and path verification of the intelligent path optimization decision module.

[0034] 3. Intelligent Path Optimization Decision Module This module is the core decision-making unit of the system. It adopts an innovative three-stage training architecture reinforcement learning algorithm and combines it with a digital twin simulation environment to achieve intelligent optimization of low-altitude flight paths.

[0035] (1) Three-stage training architecture: ①Perception pre-training stage: The perception model is trained through supervised learning to identify key elements in low-altitude flight scenarios (static obstacles, dynamic obstacles, weather areas, take-off and landing points) and establish an accurate understanding of the environment. ②Imitation learning initialization phase: Using large-scale real flight expert data (safe flight trajectories, emergency response cases) for imitation learning, the action probability distribution of the path optimization model is initialized to avoid the cold start problem of reinforcement learning training and ensure the safety and rationality of the model's initial strategy; ③ Reinforcement learning fine-tuning stage: Deploy the initialized model into the digital twin simulation environment, and perform reinforcement learning fine-tuning through real-time interaction between the agent and the simulation environment. Use a safety-oriented reward function to guide the model to learn the causal relationships of the physical world and optimize the path strategy.

[0036] (2) Safety-oriented reward function design: The reward function takes into account flight safety, mission efficiency, energy consumption cost and trajectory deviation. The expression is: R1=α·R2+β·R3-γ·R4-δ·R5. Among them, R2 is the safety reward item. When the distance between the aircraft and the obstacle is greater than the safety threshold (set according to the size of the aircraft, such as 5m for small drones), R2=5; when a collision risk is detected (distance is less than the safety threshold but greater than the warning threshold), R2=-20; when a collision occurs, R2=-100; R3 is the task efficiency reward item, R3=10-k·t (t is the current task time, k is the efficiency coefficient); R4 is the energy cost penalty item, R4=0.1·E (E is the current flight energy consumption); R5 is the trajectory deviation penalty item, R5=d (d is the deviation distance between the actual trajectory and the preset reference trajectory); α, β, γ, δ are weight adjustment parameters, with default values ​​of 0.4, 0.3, 0.2, and 0.1 respectively, which can be dynamically adjusted according to the task type (e.g., increasing the β value to improve efficiency priority for emergency rescue tasks).

[0037] (3) Multi-objective optimization constraint processing: Integrate multi-objective optimization constraint units to clarify the constraints of path optimization, including minimum safe altitude constraints (set according to terrain and building height, not less than 10m), maximum flight speed constraints (set according to airspace control requirements, not exceeding 50km / h in urban areas), energy consumption threshold constraints (ensuring the aircraft can complete the mission and return to the take-off and landing point, with remaining battery power not less than 20%), and mission time window constraints (setting the completion time limit according to mission requirements). A hierarchical constraint processing mechanism is adopted, with safety-related constraints (minimum safe altitude, no collision) as primary constraints, which are given priority; efficiency and energy consumption-related constraints are treated as secondary constraints, and optimization is carried out under the premise of satisfying the primary constraints, coordinating the conflicts between various constraints.

[0038] 4. Multi-aircraft collaborative decision-making module For multi-aircraft cooperative flight scenarios, this module is based on a distributed federated reinforcement learning mechanism to achieve cooperative optimization of multi-aircraft paths and conflict avoidance, while protecting the mission privacy of each aircraft.

[0039] (1) Distributed node deployment: Each aircraft acts as an independent local training node, and completes the training of the local path optimization model based on its own collected perception data and digital twin simulation data, without disclosing the original mission data and location information to the outside world, thus ensuring data privacy.

[0040] (2) Credibility weighted aggregation: A credibility assessment module is set up to calculate the model credibility of each local node from two dimensions: training data quality (data integrity and accuracy) and model prediction accuracy (safety and efficiency of path planning). A weighted aggregation algorithm is used to aggregate model parameters according to the credibility weight of each local model to generate a global collaborative optimization model, ensuring the accuracy and reliability of the global model.

[0041] (3) Conflict avoidance mechanism: Based on the global collaborative optimization model, the flight trajectory of each aircraft is predicted, and the possibility of trajectory intersection is evaluated by using spatial grid mapping and density function estimation methods. When potential path conflicts are detected, the flight speed, altitude or local path of the aircraft is dynamically adjusted to generate a conflict-free collaborative flight path. If the conflict cannot be resolved by local adjustment, the global path replanning process is triggered to ensure the safety of multi-aircraft collaborative flight.

[0042] 5. Virtual-Real Linkage Control Feedback Module This module enables real-time linkage between the physical flight process and the digital twin model, and dynamically corrects the path optimization model through a feedback mechanism to improve the reliability of path planning.

[0043] (1) Deviation detection: Real-time acquisition of actual flight status data (position, speed, attitude) of the aircraft, comparison with the predicted flight trajectory in the digital twin model, calculation of deviation values ​​(position deviation, speed deviation, attitude deviation); setting deviation thresholds (position deviation default 0.5m, 1m in complex environments), when the deviation value exceeds the threshold, it is determined as trajectory deviation, triggering the path adjustment process.

[0044] (2) Path dynamic adjustment: Based on the incremental learning algorithm, real-time deviation data and the latest environmental data (such as sudden obstacles and weather changes) in the digital twin model are integrated to make local fine adjustments to the current optimized path, generate the adjusted path command, and send it to the aircraft control system; at the same time, the deviation data and the adjusted path data are fed back to the intelligent path optimization decision module for continuous optimization of the model and to improve the model's adaptability to complex environments.

[0045] 6. Spatiotemporal Evolution Risk Assessment Module This module predicts the risk evolution trend of low-altitude airspace based on historical and real-time data, providing risk-oriented basis for path optimization decisions.

[0046] (1) Risk assessment model construction: Based on Bayesian network, a spatiotemporal evolution risk prediction model is constructed. Static risk factors such as terrain complexity and building density, and dynamic risk factors such as meteorological change probability, dynamic obstacle occurrence probability, and communication interruption probability are used as input nodes, and flight risk level (low risk, medium risk, high risk, and extremely high risk) is used as output node.

[0047] (2) Risk level map generation: The low-altitude flight area is divided into several grid units by using a grid division method. The risk level of each grid unit in the future preset time period (such as 5 min, 10 min) is calculated by the risk prediction model to generate a dynamic risk level map. High-risk and extremely high-risk areas are highlighted and the sources of risk (such as strong winds, dense crowds) are marked to provide a basis for the intelligent path optimization decision module to avoid them. Example

[0048] Example 1: Single-aircraft urban low-altitude logistics delivery scenario This embodiment focuses on urban low-altitude logistics delivery scenarios, realizing intelligent path optimization for a single logistics drone. The specific implementation steps are as follows: 1. System Deployment: In the target delivery area (e.g., the central urban area of ​​a city, covering an area of ​​approximately 50km²) 2 A multi-source data sensing network is deployed, including 10 fixed meteorological monitoring stations, 20 video surveillance cameras, and 5 radar detection units; logistics drones are equipped with IMU, GNSS, visual inertial navigation and communication link monitoring units; and low-altitude digital twin modeling servers and intelligent path optimization decision servers are deployed at regional edge nodes.

[0049] 2. Data Acquisition and Fusion: The multi-source data sensing and fusion module acquires static data (topographic elevation data and building outline data of the target area with a precision of 10m×10m, location and carrying capacity data of 5 delivery take-off and landing points, and airspace control boundary data) and dynamic data (real-time UAV status data: position accuracy ±0.1m, speed ±0.05m / s, real-time meteorological data: wind speed 0-15m / s, wind direction 0-360°, visibility ≥100m, dynamic obstacle data: vehicle, pedestrian, and other logistics UAV trajectories, and communication link data: bandwidth ≥10Mbps, latency ≤50ms). The weights of each sensor data are adjusted by an adaptive weighting algorithm based on entropy weight assignment. After wavelet transform denoising, a collaborative Kalman filter algorithm is used to complete data fusion and output unified environmental and UAV status data.

[0050] 3. Digital Twin Model Construction and Update: The low-altitude digital twin modeling module is based on BeiDou grid spatial coding technology. It uses a 10m×10m conventional grid to construct a digital twin model of the target area and integrates 3DGS technology to achieve detailed rendering of buildings, roads, and take-off and landing points. It integrates real-time meteorological data and dynamic obstacle data and updates the model in 1-second time steps. When a sudden obstacle is detected (such as a temporary construction area), an emergency update process is triggered, and the relevant area model is updated within 0.2 seconds.

[0051] 4. Path Optimization Decision: The intelligent path optimization decision module loads a pre-trained three-stage model (the perception pre-training stage identifies key elements of the delivery scenario, the imitation learning stage initializes the model based on 1000 historical safe delivery trajectories, and the reinforcement learning stage completes 100,000 iterations of training in a digital twin simulation environment); it inputs the delivery task requirements (starting point: delivery center A, ending point: community delivery point B, task time window: 30 min, load: 5 kg) and the fused environmental data; based on a safety-oriented reward function (α=0.4, β=0.3, γ=0.2, δ=0.1), it completes path optimization in the digital twin simulation environment and outputs the initial optimized path (length approximately 8 km, estimated time 12 min, remaining battery 35%), avoiding high-risk areas (such as densely populated commercial areas and school areas).

[0052] 5. Virtual-Real Linkage and Path Adjustment: The UAV flies along the initial optimized path. The virtual-real linkage control feedback module collects the actual flight data of the UAV in real time and compares it with the predicted trajectory of the digital twin model. When the UAV flies to an intersection, it detects a sudden vehicle entering the airspace. The deviation between the actual trajectory and the predicted trajectory reaches 1.2m (exceeding the threshold of 1m), triggering the path adjustment process. Based on the incremental learning algorithm, the sudden obstacle data is integrated to make local fine adjustments to the path (detour of 50m), generate the adjusted path command and send it to the UAV. At the same time, the deviation data is fed back to the optimization model to complete the model fine-tuning.

[0053] 6. Mission Completed: The drone safely arrived at delivery point B along the adjusted path. The mission took 13 minutes, with 32% battery remaining. There was no risk of collision, and the mission completion efficiency and safety both met the requirements.

[0054] Example 2: Multi-aircraft power line inspection scenario This embodiment focuses on the inspection scenario of high-voltage transmission lines, and realizes the collaborative path optimization of three inspection drones. The specific implementation steps are as follows: 1. System Deployment: In the target inspection area (a high-voltage transmission line in a mountainous area, approximately 20km long, covering an area of ​​approximately 80km²), 2 A multi-source data sensing network was deployed, including 5 meteorological monitoring stations, 15 video surveillance cameras, and 3 radar detection units; 3 inspection drones were equipped with IMU, GNSS, visual inertial navigation, infrared thermal imaging sensors and communication link monitoring units; and distributed federated learning servers, low-altitude digital twin modeling servers and collaborative decision-making servers were deployed at the regional edge nodes.

[0055] 2. Data Acquisition and Fusion: The multi-source data sensing and fusion module acquires static data (1m×1m high-precision terrain elevation data of the target area, power transmission line tower location and structure data, inspection take-off and landing point data, and airspace control boundary data) and dynamic data (real-time status data of 3 UAVs, real-time meteorological data: wind speed 0-20m / s, wind direction, and rainfall, dynamic obstacle data: birds, mountain wildlife, temporary workers, and communication link data); the data is fused through an adaptive weighted fusion algorithm and a cooperative Kalman filter algorithm to output high-precision environmental and UAV status data.

[0056] 3. Digital Twin Model Construction and Update: A digital twin model is constructed using a 1m×1m high-precision grid to accurately render static elements such as transmission lines, towers, and terrain; real-time meteorological data and dynamic obstacle data are integrated, and the model is updated in 0.5s time steps; when strong winds (wind speed ≥15m / s) are detected, an emergency update process is triggered.

[0057] 4. Collaborative Path Optimization Decision: Three UAVs serve as local training nodes, completing local model training based on their own perception data and digital twin simulation data; the credibility assessment module calculates the credibility of each node's model (based on data integrity and prediction accuracy, the credibility of the three UAVs are 0.35, 0.32, and 0.33, respectively); a global collaborative optimization model is generated by aggregating according to credibility weights; the inspection task requirements are input (the three UAVs are responsible for inspecting different road sections, with a total task time window of 60 minutes, requiring completion of infrared temperature measurement and visual inspection of the route); the global model completes collaborative path optimization in the digital twin simulation environment, outputting the collaborative flight paths of the three UAVs, ensuring that the paths do not intersect or conflict, and that the inspection covers all target routes.

[0058] 5. Flight and Cooperative Adjustment: The UAVs fly along a cooperative path, and the multi-UAV cooperative decision-making module monitors the trajectory of each UAV in real time. When UAV No. 2 detects a gathering of birds near a certain tower (a sudden dynamic obstacle), it generates a local avoidance path and shares the relevant information with other nodes through the federated learning network. Based on the shared information, the global model fine-tunes the flight speed of UAVs No. 1 and No. 3 to ensure the rhythm of cooperative inspection and avoid path conflicts. The virtual-real linkage control feedback module corrects the trajectory deviation of each UAV in real time to ensure inspection accuracy.

[0059] 6. Mission Completion: The three drones worked together to complete the inspection mission, which took a total of 58 minutes. There was no risk of collision, the inspection data was complete, and two minor defects in the lines were successfully identified. The mission completion quality met the requirements.

[0060] Example 3: Low-altitude flight scenario for emergency rescue This embodiment focuses on post-earthquake emergency rescue scenarios, optimizing the flight paths of rescue drones. Prioritizing both flight safety and rescue efficiency, the specific implementation steps are as follows: 1. System Deployment: In the earthquake relief area (approximately 30km²) 2 A temporary multi-source data sensing network was deployed (including collapsed buildings and temporary rescue points), consisting of 8 mobile meteorological monitoring stations, 12 video surveillance cameras, and 4 radar detection units; rescue drones were equipped with IMU, GNSS, visual inertial navigation, emergency communication modules, and life detection sensors; and a low-altitude digital twin modeling server and an intelligent path optimization decision server were deployed at the rescue command center.

[0061] 2. Data Acquisition and Fusion: The multi-source data sensing and fusion module acquires static data (post-disaster 1m×1m high-precision terrain data, collapsed building outline data, and temporary rescue point location data) and dynamic data (real-time UAV status data, real-time meteorological data: wind speed 0-18m / s, visibility ≥50m, aftershock monitoring data, dynamic obstacle data: rescue vehicles, personnel, and fallen objects, and emergency communication link data); and completes data fusion by prioritizing the weights of life detection sensors and emergency communication data through an adaptive weighted fusion algorithm.

[0062] 3. Digital Twin Model Construction and Update: A digital twin model is constructed using a 1m×1m high-precision grid to accurately render key elements such as collapsed buildings and rescue channels; real-time aftershock data, meteorological data, and dynamic rescue personnel / vehicle data are integrated, and the model is updated urgently in 0.2s time steps to ensure that the model can reflect the complex post-disaster environment in real time.

[0063] 4. Path Optimization Decision: The intelligent path optimization decision module adjusts the reward function weights (α=0.5, β=0.3, γ=0.1, δ=0.1) to prioritize safety and efficiency; input rescue mission requirements (starting point: temporary rescue command center, ending point: suspected trapped personnel area, mission: life detection and material delivery, time window: 20min); based on the dynamic risk level map generated by the spatiotemporal evolution risk assessment module (avoiding high-risk aftershock areas and collapsed material falling areas), complete path optimization in the digital twin simulation environment, and output the initial optimized path (length approximately 5km, estimated time 8min).

[0064] 5. Virtual-Real Linkage and Emergency Adjustment: The UAV flies along the initial path. During the flight, a sudden aftershock is detected, causing a secondary collapse of a local building (dynamic obstacle). The virtual-real linkage control feedback module immediately triggers the emergency path replanning process. Based on real-time updated data from the digital twin model, the replanning is completed within 1 second, generating a detour path (increased in length by 0.8km, with an estimated increase in travel time of 2 minutes). The UAV safely arrives at the target area along the new path, completing life detection and material delivery.

[0065] 6. Mission accomplished: The drone successfully completed the rescue mission in 9 minutes, with no risk of collision. It successfully located the three trapped individuals and delivered supplies precisely, providing crucial support for ground rescue efforts.

[0066] Comparison Example Compare with Example 1: Low-altitude path planning system based on traditional genetic algorithm The existing low-altitude path planning system based on genetic algorithms was adopted, but without the introduction of digital twin models and reinforcement learning algorithms. Data fusion was performed using a simple weighted average method. Testing was conducted in the urban logistics delivery scenario of Example 1, and the results are as follows: the path planning convergence speed was slow (approximately 30 seconds, compared to only 5 seconds in this invention); it could not respond to sudden obstacles in real time, and when encountering temporary construction areas, the path was not adjusted in time, leading to increased flight risk; the task completion time was 18 minutes (13 minutes in this invention), and the remaining battery power was 28% (32% in this invention), indicating lower efficiency and economy compared to this invention.

[0067] Compare with Example 2: Low-altitude path planning system based on single reinforcement learning Using a low-altitude path planning system based on single reinforcement learning in existing technologies, without imitation learning initialization and digital twin simulation training, a test was conducted in the power line inspection scenario of Example 2. The results are as follows: the model training cold start problem is serious, with a collision rate of 35% in the initial training stage (the initial collision rate of the three-stage training of this invention is only 5%); in complex dynamic environments (strong winds and bird gatherings), the path optimization adaptability is poor, with two instances of excessive trajectory deviation (this invention has no excessive deviation); the multi-aircraft collaborative decision-making adopts a centralized mode, with a data transmission delay of 200ms (the distributed mode of this invention has a delay of ≤50ms), resulting in low collaborative efficiency and a total inspection time of 85 minutes (this invention takes 58 minutes).

[0068] Compare with Example 3: A digital twin path planning system without virtual-physical linkage A path planning system with only digital twin modeling capabilities but no virtual-real linkage feedback was tested in the emergency rescue scenario of Example 3. The results are as follows: The digital twin model lags behind the physical world (update step size 5s). When a secondary building collapse occurs, the model is not updated in time, causing the planned path to pass through a dangerous area. There is no real-time deviation correction mechanism, and the actual trajectory of the drone deviates from the planned path by up to 2.5m, posing a collision risk. The task took 15 minutes to complete (compared to 9 minutes in this invention), and the material delivery was not completed in time, resulting in a lower rescue efficiency than this invention.

[0069] The above embodiments are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the invention should fall within the protection scope of the present invention. For example, adjusting the algorithm parameters of each module, expanding the types of sensors for data sensing, and adapting to different types of low-altitude aircraft (such as light helicopters and electric vertical take-off and landing aircraft) all fall within the protection scope of the present invention.

Claims

1. A flight path intelligent optimization decision support system based on a low-altitude digital twin model, characterized in that, include: The multi-source data sensing and fusion module is used to collect static and dynamic data of low-altitude flight areas and complete data preprocessing and spatiotemporal synchronization through an adaptive weighted fusion algorithm; The low-altitude digital twin modeling module constructs a digital twin of the low-altitude airspace based on multi-scale dynamic modeling technology, realizing real-time mapping and dynamic updating of the physical airspace and the virtual model; the intelligent path optimization decision module adopts a reinforcement learning algorithm with a three-stage training architecture, combined with the digital twin simulation environment to complete the training and inference of the path optimization model, and outputs the initial optimized path. The multi-vehicle collaborative decision-making module, based on a credibility-weighted distributed federated reinforcement learning mechanism, achieves collaborative optimization of multi-vehicle paths and conflict avoidance. The virtual-real linkage control feedback module collects the aircraft's flight status data in real time, feeds it back to the digital twin model for deviation correction, and drives the path optimization model to dynamically adjust path parameters. The low-altitude digital twin modeling module uses BeiDou grid spatial coding technology to achieve airspace grid management, integrates 3DGS technology to improve model rendering accuracy, and adopts an accuracy adaptive adjustment strategy to dynamically match modeling accuracy according to flight mission requirements. The three-stage training architecture includes a perception pre-training stage, an imitation learning initialization stage, and a reinforcement learning fine-tuning stage. The model is guided to learn the causal relationships of the physical world through a safety-oriented reward function, which comprehensively considers collision risk, trajectory deviation, energy consumption cost, and task completion efficiency.

2. The intelligent flight path optimization decision support system based on a low-altitude digital twin model according to claim 1, characterized in that: The static data collected by the multi-source data sensing and fusion module includes terrain elevation data, building outline data, take-off and landing point distribution data, and airspace control boundary data; the dynamic data includes real-time meteorological data, aircraft status data, dynamic obstacle data, and communication link status data; the adaptive weighted fusion algorithm dynamically adjusts the weights of each sensor data based on the entropy weight assignment method, and combines the collaborative Kalman filter algorithm to eliminate data noise and redundancy.

3. The intelligent flight path optimization decision support system based on a low-altitude digital twin model according to claim 1, characterized in that: The dynamic update mechanism of the low-altitude digital twin modeling module includes: real-time processing of multi-source sensing data based on edge computing nodes, updating the environmental parameters of the digital twin model according to a preset time step; when a sudden dynamic obstacle or meteorological change is detected, an emergency update process is triggered, prioritizing the updating of model data for key areas that affect flight safety.

4. The intelligent flight path optimization decision support system based on a low-altitude digital twin model according to claim 1, characterized in that: The safety-oriented reward function is expressed as: R1=α·R2+β·R3-γ·R4-δ·R5; where R2 is the safety reward, which is a positive reward when there is no collision risk and a negative penalty when a collision risk is detected; R3 is the task efficiency reward, which is negatively correlated with the task completion time; R4 is the energy cost penalty, which is positively correlated with flight energy consumption; R5 is the trajectory deviation penalty, which is positively correlated with the deviation distance of the preset reference trajectory; α, β, γ, and δ are weight adjustment parameters that satisfy α+β+γ+δ=1.

5. The intelligent flight path optimization decision support system based on a low-altitude digital twin model according to claim 1, characterized in that: The distributed federated reinforcement learning mechanism includes: each aircraft acts as a local training node, completing local model training based on its own perception data and digital twin simulation data; the credibility of each local node's model is calculated through a credibility evaluation module, and the credibility is determined based on the quality of training data and the accuracy of model prediction; the parameters of each local model are aggregated according to the credibility weight to generate a global optimization model, thereby achieving multi-aircraft collaborative decision-making while protecting data privacy.

6. The intelligent flight path optimization decision support system based on a low-altitude digital twin model according to claim 1, characterized in that: The virtual-real linkage control feedback module includes a deviation detection unit and a path adjustment unit. The deviation detection unit calculates the deviation between the actual flight trajectory of the aircraft and the trajectory predicted by the digital twin model. When the deviation exceeds a preset threshold, the path adjustment unit is triggered. The path adjustment unit, based on an incremental learning algorithm, integrates real-time deviation data and environmental change data to perform local fine-tuning of the optimized path.

7. The intelligent flight path optimization decision support system based on a low-altitude digital twin model according to claim 1, characterized in that: It also includes a spatiotemporal evolution risk assessment module. This module constructs a risk prediction model based on a Bayesian network, integrates historical flight data, real-time environmental data, and aircraft status data, predicts the risk evolution trend of low-altitude airspace within a preset time period, and outputs a risk level map to provide risk-oriented basis for path optimization decisions.

8. The intelligent flight path optimization decision support system based on a low-altitude digital twin model according to claim 7, characterized in that: The risk level map adopts a grid-based division method, and the risk level of each grid unit is determined by a combination of static and dynamic risk factors. Static risk factors include terrain complexity and building density; dynamic risk factors include the probability of sudden weather changes, the probability of dynamic obstacles, and the probability of communication interruption.

9. The intelligent flight path optimization decision support system based on a low-altitude digital twin model according to claim 1, characterized in that: The intelligent path optimization decision-making module also includes a multi-objective optimization constraint unit, which includes minimum safe altitude constraint, maximum flight speed constraint, energy consumption threshold constraint, and mission time window constraint. The conflict between the various constraints is coordinated through a hierarchical constraint processing mechanism.

10. The intelligent flight path optimization decision support system based on a low-altitude digital twin model according to claim 1, characterized in that: The low-altitude digital twin modeling module also integrates a scene simulation unit, which can simulate different weather conditions, obstacle distributions, and aircraft malfunction scenarios, providing a diverse simulation environment for the training and verification of the path optimization model.

Images