Unmanned aerial vehicle path planning method based on low earth orbit satellite network

Abstract

The invention provides an unmanned aerial vehicle path planning method based on a low-orbit satellite network, and the method comprises the steps: obtaining the wind field data of the low-orbit satellite network at the current position of an unmanned aerial vehicle, processing the wind field sequence data through employing a ConvLSTM network model, and extracting the wind direction trend characteristics and turbulence intensity; point cloud data sampled by a laser radar in an unmanned aerial vehicle are acquired and preprocessed, and then are used as input of a PointNet + + network model, and local features and environment features are extracted; adopting a path planning reasoning model to carry out strategy reasoning on the fused joint features, and generating an action instruction in real time; and an action instruction is sent to the flight control system through the unmanned aerial vehicle ROS system, and the action instruction is executed. Interference factors such as the wind direction in the current environment are analyzed and predicted in real time through the low-orbit satellite network, the flight path of the unmanned aerial vehicle is planned, adjusted and optimized in real time, the flight energy consumption of the unmanned aerial vehicle is reduced, and obstacles are avoided in time.

Description

Technical Field

[0001] The present application relates to the field of low-orbit satellite communication technology, and in particular to a method for unmanned aerial vehicle (UAV) path planning based on a low-orbit satellite network. Background Art

[0002] As a new type of aircraft, drones are increasingly being used in aerial photography, agriculture, plant protection, micro-selfies, express delivery, disaster relief, wildlife observation, infectious disease monitoring, surveying and mapping, news reporting, power inspections, disaster relief, and filming, playing an irreplaceable role in our lives. Low-orbit satellite communication networks provide drones with a broader range of environmental information and ensure more reliable communication. Complex external conditions pose significant challenges to drone flight, especially dynamic environmental changes, which significantly impact drone path planning. For example, flying against wind in rainy weather can significantly deplete drone energy and compromise flight path efficiency and safety. Therefore, real-time drone flight path planning to mitigate dynamic environmental challenges, as well as ensuring proper obstacle avoidance and minimizing energy consumption in these challenging environments, are key challenges in drone flight. Summary of the Invention

[0003] Given that drones face challenges such as energy efficiency in extreme environments, the present invention provides a drone path planning method based on a low-orbit satellite network. The low-orbit satellite network is used to analyze and predict interference factors such as wind direction in the current environment in real time, and to plan and adjust the drone flight path in real time, thereby reducing drone flight energy consumption, avoiding obstacles in a timely manner, and optimizing the flight path.

[0004] In a first aspect, the present invention provides a UAV path planning method based on a low-orbit satellite network, comprising: obtaining wind field data at the current UAV position from the low-orbit satellite network, generating a three-dimensional wind speed vector, and processing the wind field sequence data using a ConvLSTM network model to extract wind direction trend characteristics and turbulence intensity to obtain high-dimensional feature data 1; obtaining point cloud data sampled by a lidar in the UAV and preprocessing it as input to a PointNet++ network model to extract local features and environmental features to obtain high-dimensional feature data 2;

[0005] After fusing the high-dimensional feature data 1 and the high-dimensional feature data 2 and performing time series feature extraction, the obtained feature data is used as the input of the path planning reasoning model, and strategy reasoning is performed to generate action instructions in real time; the action instructions are sent to the flight control system through the drone ROS system to execute the action instructions; wherein, the path planning reasoning model uses the training data as the input of the PPO path planning algorithm framework, adopts a multi-stage training strategy, and is deployed in the drone system.

[0006] Furthermore, the multi-stage training strategy includes a training strategy for static obstacles and constant wind fields, a training strategy for dynamic obstacles and randomly changing wind fields, and a training strategy for multi-UAV collaboration and extreme weather.

[0007] Furthermore, the feature extraction of the point cloud data includes the following:

[0008] After the laser radar is used to sample the point cloud data, it is downsampled to 10 per frame. 4 points, retaining key structural features; using statistical outlier removal to eliminate sensor noise;

[0009] Build a PointNet++ network model and train it on a large amount of environmental data to obtain the trained PointNet++ model.

[0010] The sampled point cloud data is used as the input of the PointNet++ model to extract semantic features such as obstacles and traffic areas, and output a 256-dimensional feature vector.

[0011] Furthermore, the high-dimensional feature fusion is to fuse the wind field sequence features and the environmental point cloud data features, design the Cross-Attention module, and calculate the association weights between the point cloud features and the wind field features:

[0012]

[0013] Q is the point cloud feature; K and V are the wind field features; the output is the fused 384-dimensional feature vector.

[0014] Furthermore, the path planning reasoning model adopts the PPO network framework to construct the Actor network and the Critic network, takes the acquired environmental characteristics and wind field changes and other time series characteristics as the input of the model, and designs the reward function: R t =αR path +βR obstacle +βR energy +δR wind ;

[0015] Among them, the path efficiency reward R path , the distance reduction rate to the target point: Δd=(d t-1 -d t ) / d max , and adopts curve smoothing penalty and the sum of the absolute value of the path curvature: ∑|κ t |;

[0016] Obstacle avoidance safety reward R obstacle In the example, the laser radar detects the nearest obstacle at a distance d obs :

[0017]

[0018] The power consumption model in energy efficiency is set as P = k1υ 3 +k2ω 2 , v is the velocity, ω is the angular velocity;

[0019] Wind Farm Reward R wind , the tailwind acceleration gain is set to η·(v·ω wind ), η is the gain coefficient;

[0020] The crosswind stability penalty is set to -λ·‖ω cross ‖ 2 ,ω cross is the lateral wind component.

[0021] Furthermore, the temporal feature extraction adopts an LSTM model, and the fused features after fusing the high-dimensional feature data 1 and the high-dimensional feature data 2 are used as the input of the LSTM model, and the fused features of several frames are used as the input step size to learn dynamic environmental changes; the fused high-dimensional joint feature vector is encoded through the LSTM temporal encoding, with 5 frames of fused features as a group to learn environmental change characteristics; finally, a 256-dimensional temporal state code is output.

[0022] Through the above embodiment, there are the following advantages or beneficial effects:

[0023] (1) Using reinforcement learning algorithms to plan flight paths in real time to cope with dynamic environmental changes;

[0024] (2) Based on the low-orbit satellite network, it provides a reliable weather environment perception and prediction system to ensure low-energy flight path planning and reliable communication conditions for drones.

[0025] Other features and advantages of the present application will be described in the following description, and in part will become apparent from the description, or will be understood by practicing the present application. The purposes and other advantages of the present application can be realized and obtained by the structures particularly pointed out in the written description, claims, and drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] In order to more clearly illustrate the technical solutions in the embodiments of the present application or related technologies, the following briefly introduces the drawings required for use in the embodiments or related technical descriptions. Obviously, the drawings described below are only embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying any creative work.

[0027] Figure 1 This is a flow chart of a UAV path planning method based on a low-orbit satellite network provided by the present invention. DETAILED DESCRIPTION

[0028] To make the objectives, technical solutions, and advantages of this application more clearly understood, the technical solutions in the embodiments of this application will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of this application. Obviously, the described embodiments are only part of the embodiments of this application, not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0029] Low-Earth Orbit Satellite Network Description: The Low-Earth Orbit Satellite Network consists of a large number of small communications satellites, orbiting at an altitude of approximately 500-1200 kilometers. Compared to traditional communications satellites, it offers advantages such as wide coverage, low transmission latency (20-40 milliseconds), and high communication speeds (over 100 Mbps). It can provide high-speed and stable communication services for drones worldwide. Drones based on the Low-Earth Orbit Satellite Network can communicate with the Low-Earth Orbit Satellite Network.

[0030] Example 1

[0031] like Figure 1 As shown, a UAV path planning method based on a low-orbit satellite network includes:

[0032] Step S01: Obtain wind field data at the current UAV location from the low-orbit satellite network, generate a three-dimensional wind speed vector, and use the ConvLSTM network model to process the wind field sequence data, extract wind direction trend characteristics and turbulence intensity, and obtain high-dimensional feature data one.

[0033] Align the wind field grid data provided by the satellite with the current positioning position of the UAV to obtain the wind field data around the UAV, and generate the three-dimensional wind speed vector at the UAV's location through interpolation;

[0034] Construct a ConvLSTM network to learn the collected wind field sequence data for a period of time, extract the wind direction trend and turbulence intensity, and output a 128-dimensional feature vector, including features such as the tailwind gain system and the crosswind disturbance risk.

[0035] Step S02: Obtain and pre-process the point cloud data sampled by the UAV's laser radar, and use it as the input of the PointNet++ network model to extract local features and environmental features to obtain high-dimensional feature data 2;

[0036] The 3D point cloud data collected by the LiDAR constructs the surrounding environment. The feature extraction of the point cloud data includes the following:

[0037] After the laser radar is used to sample the point cloud data, it is downsampled to 10 per frame. 4points, reducing useless information and retaining key structural features; statistical outlier removal is used to eliminate sensor noise, such as interference points such as rain and fog.

[0038] Build a PointNet++ network model and train it on a large amount of environmental data to obtain the trained PointNet++ model.

[0039] The sampled point cloud data is used as the input of the PointNet++ model to extract semantic features such as obstacles and traffic areas, and output a 256-dimensional feature vector.

[0040] Step S03: The path planning inference model is obtained by using the training data as the input of the PPO path planning algorithm framework, adopting a multi-stage training strategy, and deployed into the UAV system;

[0041] Using the PPO network framework, we build the Actor network and the Critic network, take the acquired environmental characteristics and wind field changes and other time series characteristics as the input of the model, and design the reward function.

[0042] R t =αR path +βR obstacle +γR energy +δR wind

[0043] Among them, the path efficiency reward R path , the distance reduction rate to the target point: Δd=(d t-1 -d t ) / d max , and adopts curve smoothing penalty and the sum of the absolute value of the path curvature: ∑|κ t |.

[0044] Obstacle avoidance safety reward R obstacle In the example, the laser radar detects the nearest obstacle at a distance d obs :

[0045]

[0046] The power consumption model in energy efficiency is set as P = k1υ 3 +k2ω 2 , υ is the velocity, ω is the angular velocity. Wind field reward R wind , the tailwind acceleration gain is set to η·(υ·ω wind ), η is the gain coefficient;

[0047] The crosswind stability penalty is set to -λ·‖ω cross ‖ 2 ,ω cross is the lateral wind component.

[0048] After building the simulation environment, a multi-stage training strategy is adopted, including training strategies for static obstacles and constant wind fields, training strategies for dynamic obstacles and randomly mutated wind fields, and training strategies for multi-UAV collaboration and extreme weather.

[0049] In the first stage, the basic obstacle avoidance strategies and downwind flight environment are trained; the second stage is to increase the complexity of the environment, setting up some moving obstacles and randomly changing wind fields; the third stage adds multi-UAV collaboration scenarios and the model's adaptability in extreme weather conditions.

[0050] After multi-stage scenario training, the path planning reasoning model is obtained. After model lightweight processing, network pruning and quantization, it is deployed to the drone's embedded platform for real-time reasoning in the drone.

[0051] Step S04: After fusing the high-dimensional feature data 1 and the high-dimensional feature data 2, and performing time series feature extraction, the obtained feature data is used as the input of the path planning reasoning model, and strategy reasoning is performed to generate action instructions in real time; the action instructions are then sent to the flight control system through the drone ROS system to execute the action instructions;

[0052] The wind field sequence features and environmental point cloud data features are integrated, and the Cross-Attention module is designed to calculate the association weights between point cloud features and wind field features:

[0053]

[0054] Q is the point cloud feature; K and V are the wind field features; the output is the fused 384-dimensional feature vector.

[0055] The temporal feature extraction adopts an LSTM model, takes the fused features obtained by fusing the high-dimensional feature data 1 and the high-dimensional feature data 2 as the input of the LSTM model, and uses the fused features of several frames as the input step size to learn dynamic environmental changes.

[0056] The fused high-dimensional joint feature vector is encoded through LSTM time series, with 5 frames of fused features as a group to learn the characteristics of environmental changes; finally, a 256-dimensional time series state code is output.

[0057] The feature values extracted in real time are used as the input of the path planning reasoning model to perform strategy reasoning and generate action instructions through the Actor network. Control instructions are sent to the flight control system through ROS to complete real-time path planning, select the path with less energy consumption, avoid obstacles in advance, and improve path flight efficiency.

[0058] Through the above scheme, the following advantages or beneficial effects are achieved:

[0059] (1) Using reinforcement learning algorithms to plan flight paths in real time to cope with dynamic environmental changes;

[0060] (2) Based on the low-orbit satellite network, it provides a reliable weather environment perception and prediction system to ensure low-energy flight path planning and reliable communication conditions for drones.

[0061] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in the present invention should be included in the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be based on the scope of protection of the claims.

Claims

1. A UAV path planning method based on a low-orbit satellite network, characterized in that: include: Obtain wind field data at the current drone location from the low-orbit satellite network, generate a three-dimensional wind speed vector, and use the ConvLSTM network model to process the wind field sequence data, extract wind direction trend characteristics and turbulence intensity, and obtain high-dimensional feature data 1; After obtaining and preprocessing the point cloud data sampled by the UAV's lidar, it is used as the input of the PointNet++ network model to extract local features and environmental features to obtain high-dimensional feature data 2; After fusing the high-dimensional feature data 1 and the high-dimensional feature data 2 and performing time series feature extraction, the obtained feature data is used as the input of the path planning reasoning model, and strategy reasoning is performed to generate action instructions in real time; the action instructions are sent to the flight control system through the drone ROS system to execute the action instructions; wherein, the path planning reasoning model uses the training data as the input of the PPO path planning algorithm framework, adopts a multi-stage training strategy, and is deployed in the drone system.

2. The UAV path planning method based on a low-orbit satellite network according to claim 1, characterized in that: The multi-stage training strategy includes a training strategy for static obstacles and constant wind fields, a training strategy for dynamic obstacles and randomly changing wind fields, and a training strategy for multi-UAV coordination and extreme weather.

3. The UAV path planning method based on a low-orbit satellite network according to claim 1, characterized in that: The feature extraction of point cloud data includes the following: After the laser radar is used to sample the point cloud data, it is downsampled to 10 per frame. 4 points, retaining key structural features; Eliminate sensor noise using statistical outlier removal; Build a PointNet++ network model and train it on a large amount of environmental data to obtain the trained PointNet++ model. The sampled point cloud data is used as the input of the PointNet++ model to extract semantic features such as obstacles and traffic areas, and output a 256-dimensional feature vector.

4. The UAV path planning method based on a low-orbit satellite network according to claim 1, wherein: The high-dimensional feature fusion is to fuse the wind field sequence features and the environmental point cloud data features, design the Cross-Attention module, and calculate the association weights between the point cloud features and the wind field features: Q is the point cloud feature; K and V are the wind field features; the output is the fused 384-dimensional feature vector.

5. The UAV path planning method based on a low-orbit satellite network according to claim 1, wherein: The path planning reasoning model adopts the PPO network framework to build the Actor network and the Critic network. The acquired environmental characteristics and wind field changes and other time series characteristics are used as the input of the model, and the reward function is designed: R t =αR path +βR obstacle +γR energy +δR wind ; Among them, the path efficiency reward R path , the distance reduction rate to the target point: Δd=(d t-1 -d t ) / d max , and adopts curve smoothing penalty and the sum of the absolute value of the path curvature: ∑|κ t |; Obstacle avoidance safety reward R obstacle In the example, the laser radar detects the nearest obstacle at a distance d obs : The power consumption model in energy efficiency is set as P = k1v 3 +k2ω 2 , υ is the velocity, ω is the angular velocity; Wind Farm Reward R wind , the tailwind acceleration gain is set to η·(υ·ω wind ), η is the gain coefficient; The crosswind stability penalty is set to -λ·‖ω cross ‖ 2 ,ω cross is the lateral wind component.

6. The UAV path planning method based on a low-orbit satellite network according to claim 1, wherein: The temporal feature extraction adopts an LSTM model, takes the fused features after fusing high-dimensional feature data 1 and high-dimensional feature data 2 as the input of the LSTM model, and uses several frames of fused features as the input step size to learn dynamic environmental changes; the fused high-dimensional joint feature vector is encoded through LSTM temporal coding, with 5 frames of fused features as a group, to learn environmental change characteristics, and finally outputs a 256-dimensional temporal state code.

Images