Autonomous Flight Decision-Making Method for Fixed-Wing UAVs Based on a Brain-Inspired Decision-Making Framework

Abstract

This invention discloses an autonomous flight decision-making method for fixed-wing unmanned aerial vehicles (UAVs) based on a brain-like decision-making framework. The method includes initializing the flight state, the starting point of the autonomous flight mission, and the target point; acquiring environmental information within the current observation window using onboard sensors, establishing a grid map, and obtaining the current flight state; encoding and compressing the grid map and current flight state within the observation window using a hybrid state encoder to obtain window observation information; using a trained neural network to output a planning step size for the window based on the observed information; performing path planning within the current observation window using a local planner to obtain sub-target points; determining whether the current sub-target point equals the target point; if so, the autonomous flight decision-making ends; otherwise, updating the window position information. This invention can solve the autonomous flight problem of high-speed fixed-wing UAVs with only local environmental perception information, can adapt to complex and dynamic mission scenarios, and features high planning efficiency and smooth path generation.

Description

Technical Field

[0001] This invention relates to the field of autonomous flight decision-making technology, and more specifically to an autonomous flight decision-making method for fixed-wing unmanned aerial vehicles based on a brain-like decision-making framework. Background Technology

[0002] When global map information between the starting point and the target point is known, path planning can be performed using pre-planning. The advantage of pre-planning is that it does not rely on onboard computing resources and can obtain the globally optimal path; the disadvantage is that it requires extensive information collection and mapping work in advance to create a map of the mission environment, and the pre-planned path often cannot respond effectively to environmental changes in a timely manner. To ensure the successful execution of missions, current technological developments increasingly focus on improving the autonomy of UAV flight.

[0003] Autonomous fixed-wing UAVs require onboard sensors to acquire information about their surroundings. Given a fixed sensing space, with the same quantization bit depth, a higher sampling rate and update rate from the sensors necessitate a larger communication bandwidth. For example, the communication bandwidth of a high-performance synthetic aperture radar in terrain mapping mode is typically several hundred Mbps. Unlike communication bandwidth, control bandwidth refers to the frequency range of signal tracking capability within a closed-loop control system. Its physical meaning reflects the system's dynamic response speed and its ability to reproduce input signals; a larger control bandwidth results in a faster system response. To meet the response requirements of rapid maneuvering and precise trajectory tracking, the servo control bandwidth of high-speed, high-maneuverability fixed-wing UAVs is typically 9-21 Hz to ensure rapid response of actuators to control commands. With upgrades to digital flight control computers, the servo bandwidth of modern high-performance fixed-wing UAVs can be even greater to support continuous 9G maneuvers. In practical engineering applications, the controller frequency used for flight control is typically 5-10 times the control bandwidth to ensure phase margin and stability. Therefore, if the autonomous flight decision unit is directly used for control surface control, it needs to possess a very high decision frequency. Meanwhile, during the autonomous flight of a fixed-wing UAV, if every signal used for control surfaces is directly output by the autonomous flight decision unit, the sampling rate and update rate of the airborne sensors should not be less than the decision frequency to ensure that new sampling data is available in each decision cycle; this will put enormous pressure on the airborne communication bandwidth and the information processing of the computing unit.

[0004] For high-speed flight scenarios, fixed-wing UAVs require a large perception space to ensure sufficient reaction time for autonomous flight decision-making and action execution, while also requiring a high control frequency to achieve rapid maneuvering and precise trajectory tracking; however, existing technologies struggle to reconcile the conflicting demands for communication and control bandwidth while ensuring decision-making efficiency. Summary of the Invention

[0005] The purpose of this invention is to propose an autonomous flight decision-making method for fixed-wing unmanned aerial vehicles (UAVs) based on a brain-inspired decision-making framework. This method is applicable to high-speed autonomous flight of fixed-wing UAVs in complex and dynamic mission scenarios. This invention abstracts the action space based on a brain-inspired decision-making framework, reconciling the conflict between communication and control bandwidth requirements while utilizing heuristic information from the underlying planner to solve training challenges under some observable states. By introducing traditional planning methods into the sequential decision-making process, this method can output a feasible trajectory and reference control variables for the UAV to track its flight over a period of time after acquiring surrounding environmental information each time.

[0006] To achieve the above objectives, the present invention provides the following technical solution: The present invention provides an autonomous flight decision-making method for fixed-wing unmanned aerial vehicles based on a brain-like decision-making framework, comprising the following steps: Step S1: Set the initial flight status and set the starting point and target point of the autonomous flight mission; Step S2: Obtain environmental information in the current observation window based on airborne sensors and build a grid map, mark the grid cell occupied by each obstacle, and obtain the current flight status; Step S3: Use a hybrid state encoder to encode and compress the grid map and current flight state obtained in step S2 to obtain the state information in the observation window, which is used as the input to the neural network; Step S4: Use the trained neural network to output the planning step size for this window based on the observation state information obtained in step S3; Step S5: Use a local planner within the current observation window. The local planner performs path planning within the local window based on the step size output by the neural network in step S4, obtains the sub-target point, and outputs the feasible spatial trajectory and corresponding reference control quantity for UAV tracking flight within the window. Step S6: Determine whether the sub-target point obtained from the window planning in step S5 is equal to the target point. If it is equal, the planning ends; otherwise, update the window position information, set the flight status of the sub-target point to the current state, and then return to step S2.

[0007] Furthermore, step S3 specifically includes: Step S31: Convert the binary grid map within the observation window into a three-dimensional matrix representation, and compress and encode the grid map; Step S32: Thirteen state variables are used to describe the flight state of the fixed-wing UAV, including position information and yaw angle information; given the known position of the target point, the horizontal distance between the current position of the UAV and the target point, the angle between the target azimuth and the UAV yaw angle, and the height difference between the current position and the target point are extracted as core position-related features. Step S33: Normalize all flight state variables to obtain a normalized set of flight states; Step S34: The encoded grid map and the normalized flight status are represented by sets to form single observation information in the observation space, which is then used as input to the neural network.

[0008] Furthermore, in step S31, the strategy for compressing and encoding the grid map is as follows:

[0009] in, , and These represent the mesh elements of the three-dimensional matrix in the length, width, and height directions, respectively. , and These represent the observed variable values ​​after compression encoding in the length, width, and height directions of the three-dimensional matrix, respectively.

[0010] Furthermore, in step S32, the formulas for calculating the horizontal distance between the current position and the target point, the angle between the target azimuth and the UAV yaw angle, and the altitude difference between the current position and the target point are as follows:

[0011] in, The horizontal distance between the current position and the target position. This is the angle between the target azimuth at the current position and the aircraft yaw angle. This represents the height difference between the current position and the target position.

[0012] Further, step S4 specifically includes: Step S41: Construct a multi-layer feedforward fully connected neural network as the interactive decision layer. The input of the neural network is a continuous observation state space, and the output is a discrete planning step size action space. Step S42: Design the action space based on the size of the grid cells used for environmental mapping, i.e., the range of values ​​for the planned step size. The minimum and maximum values ​​of the planned step size are determined based on the flight speed of the fixed-wing UAV and the mission scenario.

[0013] Furthermore, in step S41, when the input space dimension is 441 and the action space size is 10, the representation of the 4-layer neural network is as follows:

[0014] in This represents the input vector of the input layer. Represents the weight matrix. This represents the bias vector. Indicates the output of the hidden layer. This represents the activation function. This represents the output vector of the output layer; .

[0015] Furthermore, before training the neural network, a weighted reward function based on path length and planning time is designed for reinforcement learning training. The training process uses negative rewards; that is, the shorter the planned path length, the greater the reward for path length; and the less time spent on planning, the greater the reward for planning time. The reward value for each step consists of the weighted sum of these two factors, i.e.:

[0016] in, Represents the total reward value. This represents the path length reward. This indicates a time-based reward. The relative weights between the path length reward and the planning time reward are shown in the following formulas:

[0017] in, The length of the globally optimal path. Let the path length within the window be obtained using a certain step-length planning after each observation. This represents the total number of grid cells contained in the window. This represents the number of grid cells expanded during a specific window planning process.

[0018] Furthermore, when the planning reaches the target point, an additional terminal reward is added to accelerate training convergence. Let the number of planning operations in a given round be . Then, at the end of the round, the reward for that round is as follows:

[0019] in, Indicates the final value reward. , This indicates the reward for that round.

[0020] Based on the above technical solution, the embodiments of the present invention can produce at least the following technical effects: This invention addresses the resource and performance conflict in autonomous decision-making for high-speed unmanned aerial vehicles (UAVs) through a multi-dimensional collaborative mechanism. First, by projecting the environmental grid map along multiple spatial dimensions and converting the flight state into relative state features with respect to the target point, a low-dimensional observation state is constructed. This technique utilizes the principles of spatial projection and relative coordinate system transformation to significantly reduce the dimensionality of the input data while preserving the environmental topology and key mission-oriented information. This effectively reduces the computational load and communication bandwidth consumption of the onboard computing unit, making high-frequency decision-making possible. Second, by inputting the low-dimensional observation state into a pre-trained decision model and dynamically outputting the search step size, the rigid mode of traditional fixed-step-size planning is changed. Utilizing the nonlinear mapping capability of machine learning models to the current local environmental complexity and flight status, adaptive adjustment of the planning granularity is achieved. In simple environments, the step size is increased to improve planning speed; in complex environments, the step size is decreased to ensure trajectory safety and smoothness. This significantly improves the efficiency and quality of path planning while ensuring real-time control bandwidth requirements. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the structures shown in these drawings without creative effort.

[0022] Figure 1 This is a flowchart of the autonomous flight decision-making method for fixed-wing unmanned aerial vehicles proposed in this invention; Figure 2 This is a diagram of the brain-like decision-making framework for autonomous flight proposed in this invention. Figure 3 This is a schematic diagram of the observation space based on the hybrid state encoder proposed in this invention; Figure 4 This is a schematic diagram of the autonomous decision-making algorithm based on a brain-like decision-making framework proposed in this invention. Figure 5 This is the average reward curve of the autonomous decision-making algorithm during the training process in a specific embodiment; Figure 6 The specific implementation shows the planning paths of the three algorithms in a certain random environment; Figure 7 The above represents the average planning results of the three algorithms in 100 random environments in a specific embodiment. Detailed Implementation

[0023] The technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. In addition, the technical solutions of the various embodiments can be combined with each other, but this must be based on the ability of those skilled in the art to implement them. When the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such combination of technical solutions does not exist and is not within the scope of protection claimed by the present invention.

[0024] See Figure 1 , Figure 1 This is a flowchart of the autonomous flight decision-making method for fixed-wing UAVs based on a brain-like decision-making framework proposed in this invention. The fixed-wing UAV establishes a grid map based on environmental information obtained from airborne sensors such as radar and cameras, marks the grid cells occupied by each obstacle, and obtains the current flight status; during initial operation, the starting point S and target point G for path planning are set. The specific steps are as follows: Step S1: Set the initial flight state and set the starting point S and target point G for autonomous flight.

[0025] Step S2: Obtain environmental information in the current observation window based on airborne sensors and build a grid map, mark the grid cell occupied by each obstacle, and obtain the current flight status.

[0026] During autonomous flight, onboard sensors continuously perceive the surrounding environment. For example, raw point cloud or image data is collected using sensors such as lidar, millimeter-wave radar, or depth cameras, and a local environmental grid map centered on the drone's current location is constructed in real time. This environmental grid map represents the spatial distribution of obstacles within the current observation window. Simultaneously, the drone's current flight status, including but not limited to position, velocity, attitude, and angular velocity, is obtained through onboard inertial measurement units, pitot tubes, and satellite navigation systems. This step provides the initial perception data foundation for subsequent decision-making.

[0027] Step S3: Use a hybrid state encoder to encode and compress the grid map and current flight state obtained in step S2 to obtain the state information in the observation window, which is used as the input to the neural network.

[0028] Because high-resolution 3D environmental mesh maps generate massive amounts of data, directly using them for real-time decision-making would overload onboard computing resources and limit communication bandwidth. Therefore, this embodiment performs dimensionality reduction projection processing on the environmental mesh map along multiple spatial dimensions such as length, width, and height, compressing high-dimensional 3D spatial information into low-dimensional environmental spatial feature vectors. This processing method significantly reduces data dimensionality while preserving key information such as environmental topological connectivity. Simultaneously, to eliminate the impact of changes in the starting point position in the global absolute coordinate system on the decision-making strategy, the current flight state is converted into relative state features relative to the target point, such as relative distance and relative orientation. Subsequently, the environmental spatial features and relative state features are fused to construct a low-dimensional observation state that fully characterizes the current decision-making situation. This hybrid encoding strategy effectively resolves the contradiction between high-dimensional environmental perception and onboard real-time computing capabilities, and provides the decision-making model with standardized inputs that are invariant to translation and rotation, which is beneficial for improving the model's generalization ability and convergence speed.

[0029] Specifically, please refer to Figure 3 This invention uses a binary mesh map, commonly used in path planning, to model the environment. The environment mesh map is a binary 3D matrix representing obstacle occupancy; the element values ​​in the binary 3D matrix indicate whether the corresponding mesh cell is occupied by an obstacle. For example, when the physical space corresponding to a mesh cell is occupied by a mountain, building, or other aircraft, the value of that element is 1; if it is an open, flyable area, the value is 0. For a size of... The grid map in the observation window can be mapped to a size of [size missing] through a unique encoding strategy. The continuous state space. Step S3 is as follows: S31: Convert the binary grid map in the observation window into a 3D matrix. The strategy for compressing and encoding grid maps is shown in the following formula:

[0030] in, , and These represent the mesh elements of the three-dimensional matrix in the length, width, and height directions, respectively. , and These represent the observed variable values ​​after compression encoding in the length, width, and height directions of the three-dimensional matrix, respectively.

[0031] S32: Uses 13 state variables to manage the flight status of fixed-wing UAVs. Describe, in which , and These represent the airspeed, angle of attack, and sideslip angle of the UAV, respectively. , and These represent the roll angle, pitch angle, and yaw angle of the UAV, respectively. , and These represent the roll rate, pitch rate, and yaw rate of the UAV, respectively. , and This indicates the three-dimensional spatial position of the UAV in the northeast-northeast coordinate system. This indicates the actual power of the drone's engine. At the known endpoint position ( , , In the case of the aircraft's current position ( , , ) and yaw angle The processing method is shown in the following formula:

[0032] in, The horizontal distance between the current position and the target position. This is the angle between the target azimuth at the current position and the aircraft yaw angle. This represents the height difference between the current position and the target position.

[0033] S33: All flight state variables must be normalized to improve the stability of neural network training. The normalized flight states are represented in set form, as shown in the following equation:

[0034] S34: The encoded grid map and the normalized flight status are represented by a set to constitute an observation in the observation space. As shown in the following formula:

[0035] Assuming the aircraft is flying at a fixed altitude, the grid map in the observation window has only one grid cell in the altitude direction, i.e. Therefore, when processing information from only a single observation window, the observed variables can be ignored. and normalized height difference In this embodiment, the size of the observation window is set to 215×215 grid cells, that is, Each grid cell is 200m in size. After passing through the hybrid state encoder, the dimension of the observation space is 441 (after dimensionality reduction, the dimension is 215+215+11=441, and the observation space contains s). i and sj It also includes s flight The 11 normalized variable values ​​(the altitude difference was ignored due to the fixed-altitude flight) were used as input. During subsequent algorithm training, only one frame of observation information was input at a time; this was done to avoid introducing historical observation information and to verify that the autonomous decision-maker designed based on the brain-like decision-making framework possesses the ability to learn under partially observable conditions.

[0036] Step S4: Refer to Figure 2 As the "interactive decision layer" corresponding to the "brain" in the brain-like decision-making framework, the trained neural network can output the planning step size Step_size for the window based on the observation state information obtained in step S3. See Figure 4 This embodiment uses a deep Q-network to design the "brain" of the interactive decision-making layer. The input of the neural network is a continuous state space, and the output is a discrete action space. The purpose of training the neural network is to enable it to determine the step size of the underlying planner for local window planning based on the observed input. Step S4 is as follows: S41: A multi-layer feedforward fully connected neural network is used for network structure design; in this embodiment, when the input space dimension is 441 and the action space size is 10, the 4-layer neural network can be represented as follows:

[0037] in, This represents the input vector of the input layer. Represents the weight matrix. This represents the bias vector. Indicates the output of the hidden layer. This represents the activation function. This represents the output vector of the output layer; .

[0038] S42: Before training the neural network, action space design (i.e., planning the range of step size) is required. In this embodiment, the size of the grid cell is set to 200m, the expected flight speed of the fixed-wing UAV is 165m / s, and the action space design is shown in the following formula:

[0039] in, This represents the step size used for a specific window planning operation. Since fixed-wing UAVs need to consider the size of the grid cells when planning feasible paths, the minimum step size used for planning in this embodiment is... The maximum step size used for planning .

[0040] S43: Before training the neural network, a reward function is designed based on the length of feasible paths and the planning time. To learn how to find shorter paths faster, a negative reward is used for the training process. That is, the shorter the planned path, the greater the reward for path length. The larger the value, the less time is spent on planning, and the higher the planning time bonus. The larger the sum, the greater the reward. The total reward value for each step consists of the weighted sum of the two, i.e. . The value reflects the importance the algorithm's training objective places on path length and planning time. In this embodiment, shorter path lengths are prioritized, therefore the value is set to... See also Figure 5 , where is the average reward curve of the autonomous decision-making algorithm during the training process in this embodiment; path length reward and planning time rewards The following formulas are shown respectively:

[0041] in, The length of the globally optimal path. Let the path length within the window be obtained using a certain step-length planning after each observation. This represents the total number of grid cells contained in the window. The number of grid cells expanded during a window planning process is the number of grid cells expanded during each window planning process. Therefore, this invention uses the number of grid cells expanded during each window planning process as a measure of planning time, which can avoid the impact of hardware platform and system timing errors on planning time.

[0042] In addition to process rewards, to accelerate the training process, an extra final reward is added when a plan reaches its endpoint. Let the number of planning operations in a given round be . Then, at the end of this round, the reward for this round is as shown in the following formula, where, .

[0043]

[0044] Step S5: As the "cerebellum + brainstem" in the brain-like decision-making framework, the step size obtained in step S4 is input into the local planner CLRaA*_win(Step_size) to perform path planning within the local window, obtain sub-target points, generate feasible spatial trajectories and corresponding reference control quantities, and output feasible spatial trajectories for UAV tracking flight. See Figure 2 ,like Figure 2As shown, by simulating the functional hierarchy of the human brain's decision-making center, the brain-like decision-making framework proposed in this invention comprises a three-level structure: "interactive decision-making," "motion planning," and "action control." The underlying planner, CLRaA*_win, acting as a "cerebellum + brainstem," can quickly generate feasible spatial trajectories and reference control quantities for tracking these trajectories. The generation of these reference control quantities considers the control system closely related to the aircraft itself (i.e., the "brainstem"). The underlying planner of this invention references a path planning method based on a variable step size mechanism suitable for fixed-wing unmanned aerial vehicles disclosed in Chinese Patent ZL202310624432.9. Since the step size of the planner in this invention is given by the interactive decision-making layer, elliptical constraints during the path search process do not need to be considered. Because the planner operates within a local observation window in this invention, it has the ability to pause planning at sub-target points and upload the current feasible spatial trajectory and reference control quantities. During the search process, the underlying planner needs to maintain two storage lists (OPEN list and CLOSED list) and one storage set (OPEN_again set). The OPEN list stores nodes to be expanded, the CLOSED list stores nodes that have already been expanded, and the OPEN_again set stores nodes that need to be expanded with variable step size. The specific steps are as follows: S51: Initialize the CLOSE list and OPEN_again set to empty, put the starting node of the plan into the OPEN list, and obtain the planning step size k output by the interactive decision layer; S52: Determine if the OPEN list and OPEN_again are both empty. If so, output "path planning failed" and terminate the search; otherwise, proceed to step S53. S53: Select the node with the smallest evaluation function value in the OPEN list as the current node. Then, the current node is removed from the OPEN list and added to the CLOSED list; S54: Determine whether the current node is the target node or whether the current node has exceeded the current observation window. If it is the target node, obtain the planned path with a search step size of k, output the feasible spatial trajectory and reference control quantity in the current window, and end step S5; if it has exceeded the current observation window, extract the parent node of the current node as the child target node, output the feasible spatial trajectory and reference control quantity in the current window, and end step S5; otherwise, execute step S55. S55: Along the horizontal, vertical, and diagonal directions, perform collision detection on the eight neighboring nodes that are k grid cells away from the current node. For neighboring nodes that do not collide... If its path cost Smaller than the neighbor node If the previous path cost is calculated, then proceed to step 56; for the current node... After traversing all neighbor nodes, proceed to step S57; S56: For nodes Perform a feasibility check; if the node is within the reachable region, update the node. of The value will be the current node. Set as The parent node, and Add the node to the OPEN list; otherwise, continue traversing the next neighbor node. S57: Determine the current node Does the number of collisions with its neighboring nodes not less than D-1 (D is the number of expansion directions) times, or does a neighboring node exist? Path cost Less than If the previous path cost is true, then the current node will be... Add to the OPEN_again collection; S58: Check if the OPEN list is empty. If it is, then set the current node... Add the nodes to the OPEN_again set and expand all nodes in the set with variable step size. Then set the OPEN_again set to empty and return to step S52.

[0045] Specific evaluation function The formula for calculating the value is:

[0046] in, For the current node, The path cost to the current node, i.e., from the starting node to the current node. The actual cost; This is a heuristic function, representing the current node. The estimated cost to the target point can be measured using metrics such as Euclidean distance or Manhattan distance, depending on the actual application scenario. The feasibility detection in step S56 and the variable step-size expansion in step S58 are the same as the methods disclosed in Chinese Patent ZL202310624432.9.

[0047] Step S6: Determine whether the sub-target point obtained from the window planning in step S5 is equal to the target point G. If it is equal, the planning ends; otherwise, update the window position information, set the flight status of the sub-target point to the current state, and return to step S2.

[0048] After the above steps Figure 6The diagram shows the planning paths of the three algorithms in a specific embodiment under a random environment. It can be seen that the trajectory planned by Q-CLRaA* is smoother. In this specific embodiment, the number of grid cells in the planning space is 1000×1000, the grid cell size is 200 meters, and the global planning space is 200km×200km. Comparative experiments were conducted using CL_RaA*, Random-CLRaA*, and Q-CLRaA* respectively; all three algorithms can generate feasible spatial trajectories. Figure 7 This is the average planning result of the three algorithms in 100 random environments in a specific embodiment; from Figure 7 As can be seen from (a), the feasible spatial trajectory planned by the method proposed in this invention has the shortest average length and the smallest standard deviation; from Figure 7 As can be seen in (b), under the premise of ensuring the shortest planned trajectory length, the planning time of Q-CLRaA* is significantly improved compared to the CL-RaA* algorithm. This indicates that compared to Random-CLRaA* with randomly selected step size, the autonomous decision-maker has effectively learned under partially observable conditions and can find the shortest feasible spatial trajectory as quickly as possible.

[0049] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for autonomous flight decision-making of fixed-wing unmanned aerial vehicles based on a brain-like decision-making framework, characterized in that, Includes the following steps: Step S1: Set the initial flight status and set the starting point and target point of the autonomous flight mission; Step S2: Obtain environmental information in the current observation window based on airborne sensors and build a grid map, mark the grid cell occupied by each obstacle, and obtain the current flight status; Step S3: Use a hybrid state encoder to encode and compress the grid map and current flight state obtained in step S2 to obtain the state information in the observation window, which is used as the input to the neural network; Step S4: Use the trained neural network to output the planning step size for this window based on the observation state information obtained in step S3; Step S5: Use a local planner within the current observation window. The local planner performs path planning within the local window based on the step size output by the neural network in step S4, obtains the sub-target point, and outputs the feasible spatial trajectory and corresponding reference control quantity for UAV tracking flight within the window. Step S6: Determine whether the sub-target point obtained from the window planning in step S5 is equal to the target point. If it is equal, the planning ends; otherwise, update the window position information, set the flight status of the sub-target point to the current state, and then return to step S2.

2. The method for autonomous flight decision-making of fixed-wing UAVs based on a brain-like decision-making framework according to claim 1, characterized in that, Step S3 specifically includes: Step S31: Convert the binary grid map within the observation window into a three-dimensional matrix representation, and compress and encode the grid map; Step S32: Use 13 state variables to describe the flight state of the fixed-wing UAV. Given the known position of the target point, extract the horizontal distance between the current position of the UAV and the target point, the angle between the target azimuth and the UAV yaw angle, and the height difference between the current position and the target point as core position-related features. Step S33: Normalize all flight state variables to obtain a normalized set of flight states; Step S34: The encoded grid map and the normalized flight status are represented by sets to form single observation information in the observation space, which is then used as input to the neural network.

3. The method for autonomous flight decision-making of fixed-wing UAVs based on a brain-like decision-making framework according to claim 2, characterized in that, In step S31, the strategy for compressing and encoding the grid map is as follows: ; in, , and These represent the mesh elements of the three-dimensional matrix in the length, width, and height directions, respectively. , and These represent the observed variable values ​​after compression encoding in the length, width, and height directions of the three-dimensional matrix, respectively.

4. The method for autonomous flight decision-making of fixed-wing UAVs based on a brain-like decision-making framework according to claim 2, characterized in that, In step S32, the 13 state variables are adopted. Describe, in which , and These represent the airspeed, angle of attack, and sideslip angle of the UAV, respectively. , and These represent the roll angle, pitch angle, and yaw angle of the UAV, respectively. , and These represent the roll rate, pitch rate, and yaw rate of the UAV, respectively. , and This indicates the three-dimensional spatial position of the UAV in the northeast-northeast coordinate system. This indicates the actual power of the drone's engine.

5. The method for autonomous flight decision-making of fixed-wing UAVs based on a brain-like decision-making framework according to claim 4, characterized in that, In step S32, the formulas for calculating the horizontal distance between the current position and the target point, the angle between the target azimuth and the UAV yaw angle, and the height difference between the current position and the target point are as follows: ; in, The horizontal distance between the current position and the target position. This is the angle between the target azimuth at the current position and the aircraft yaw angle. This represents the height difference between the current position and the target position.

6. The method for autonomous flight decision-making of fixed-wing UAVs based on a brain-like decision-making framework according to claim 1, characterized in that, Step S4 specifically includes: Step S41: Construct a multi-layer feedforward fully connected neural network as the interactive decision layer. The input of the neural network is a continuous observation state space, and the output is a discrete planning step size action space. Step S42: Design the action space based on the size of the grid cells used for environmental mapping, i.e., the range of values ​​for the planned step size. The minimum and maximum values ​​of the planned step size are determined based on the flight speed of the fixed-wing UAV and the mission scenario.

7. The method for autonomous flight decision-making of fixed-wing UAVs based on a brain-like decision-making framework according to claim 6, characterized in that, In step S41, when the input space dimension is 441 and the action space size is 10, the representation of the 4-layer neural network is as follows: ; in, This represents the input vector of the input layer. Represents the weight matrix. This represents the bias vector. Indicates the output of the hidden layer. This represents the activation function. This represents the output vector of the output layer; .

8. The method for autonomous flight decision-making of fixed-wing UAVs based on a brain-like decision-making framework according to claim 1, characterized in that, Before training the neural network, a weighted reward function based on path length and planning time is designed for reinforcement learning training. The training process uses negative rewards: the shorter the planned path length, the greater the reward; the less time spent on planning, the greater the reward. The reward value for each step is the weighted sum of these two factors. ; in, Represents the total reward value. This represents the path length reward. This indicates a time-based reward. The relative weights between the path length reward and the planning time reward are shown in the following formulas: ; in, The length of the globally optimal path. Let the path length within the window be obtained using a certain step-length planning after each observation. This represents the total number of grid cells contained in the window. This represents the number of grid cells expanded during a specific window planning process.

9. The method for autonomous flight decision-making of fixed-wing UAVs based on a brain-like decision-making framework according to claim 8, characterized in that, When the plan reaches the target point, an additional terminal reward is added to accelerate training convergence. Let the number of planning operations in a given round be . Then, at the end of the round, the reward for that round is as follows: ; in, Indicates the final value reward. , This indicates the reward for this round.

Images