Model-data hybrid driving autonomous obstacle avoidance method for fixed-wing unmanned aerial vehicle

Abstract

This invention discloses a model-data hybrid-driven autonomous cognitive obstacle avoidance method for fixed-wing unmanned aerial vehicles (UAVs), comprising the following steps: establishing a slow-path instantaneous maneuvering obstacle avoidance model; establishing a fast-path adaptive pose control model; and establishing a brain-like fast-slow dual-path hybrid-driven maneuvering optimization model. This invention employs a brain-like fast-slow dual-path architecture, rationally designs long-term and short-term decision-making game strategies, realizes effective reinforcement mechanisms for both fast and slow paths, implements dynamic game theory between paths, and ensures that the two paths are both relatively independent and mutually reinforcing in the final decision selection. This allows the entire brain-like fast-slow dual-path model to effectively form intuitive habits through iteration and feedback, thereby improving autonomous cognitive performance. It effectively maps artificial intelligence models to human brain cognitive mechanisms, truly heuristically constructing the required autonomous intelligent system based on the interaction mechanisms and principles of the fast-slow dual-path in the brain and cognition, enabling the autonomous intelligent system to achieve intelligent cognition.

Description

Technical Field

[0001] This invention relates to autonomous cognitive obstacle avoidance technology for unmanned platforms, and in particular to a model-data hybrid-driven autonomous cognitive obstacle avoidance method for fixed-wing unmanned aerial vehicles. Background Technology

[0002] Autonomous intelligent unmanned aerial vehicle (UAV) systems combine artificial intelligence with UAV systems and have experienced rapid development in recent years due to their intelligent, unmanned, and autonomous characteristics. Fixed-wing UAVs, as a typical example of an autonomous intelligent UAV system, possess high maneuverability, high speed, and the ability to fly at high altitudes for extended periods, giving them unique advantages in military reconnaissance and target engagement, and playing an increasingly important role in the military field. Flight control is fundamental to the reliable maneuverability of fixed-wing UAVs. Faced with increasingly complex flight environments, autonomous obstacle avoidance during flight has become an essential capability for modern fixed-wing UAVs to protect their surroundings and their own safety.

[0003] Currently, the maneuvering flight control of fixed-wing UAVs is basically based on traditional control frameworks, achieving smooth operation within limited roll and pitch attitudes. However, it does not adequately address issues such as nonlinearity, strongly coupled control, Euler angle singularities, and the generation of complex maneuvers encountered during maneuvering flight. Furthermore, because fixed-wing UAVs often perform long-duration high-altitude operations, in addition to common external interferences, they must also address uncertainties arising from the wear or degradation of internal components and changes in geographical environment information—that is, the deviation between the actual dynamic model and the ideal dynamic model of the fixed-wing UAV. Especially in complex environments such as urban and mountainous areas, where space is limited, obstacles are numerous, and wind conditions are more complex, the requirements for efficiency and accuracy in real-time tracking flight are even higher. To date, autonomous obstacle avoidance technology for fixed-wing UAVs under high-dynamic conditions remains a pressing problem to be solved. Summary of the Invention

[0004] To address the aforementioned problems in existing technologies, this invention proposes a model-data hybrid-driven autonomous cognitive obstacle avoidance method for fixed-wing unmanned aerial vehicles (UAVs) that features high computational speed and accuracy.

[0005] To achieve the above objectives, the technical solution of the present invention is as follows: A model-data hybrid driven autonomous cognitive obstacle avoidance method for fixed-wing unmanned aerial vehicles, comprising the following steps:

[0006] A. Establish a slow-path instantaneous maneuvering obstacle avoidance model

[0007] First, a complex environment autonomous perception module is established to obtain the vector of the fixed-wing UAV's perception of the external environment. Then, an instantaneous prediction and timing memory module is established to predict the state of the fixed-wing UAV in a timing sequence. Finally, an attitude maneuver autonomous control module is established to generate the optimal control behavior of the fixed-wing UAV. The specific steps are as follows:

[0008] A1. Establish an autonomous perception module for complex environments.

[0009] The system observes the external environment using radar and cameras to obtain input images. These images are then compressed frame by frame, and the video frames are encoded into low-dimensional latent vectors using a variational self-encoding and decoding structure, enabling autonomous perception of complex environments.

[0010] First, the resolution of the input image is compressed to 64×64×3. Then, the image encoding is reduced to a two-dimensional space to obtain μ and σ vectors. The forms of μ and σ vectors are defined as follows:

[0011] μ = {μ1, μ2, ..., μ} d}

[0012] σ={σ1,σ2,…,σ d}

[0013] Where d represents the dimensions of the μ and σ vectors.

[0014] Let f θ f represents the encoding network for the μ vector. φ The formulas for calculating the μ vector and σ vector in the encoding network representing the σ vector are:

[0015] μ = f θ (x)

[0016] σ=f φ (x)

[0017] In the formula, x is the input image. The compression feature Z of the input image is:

[0018] Z = {z1, z2, ..., z} d}

[0019] Where: z j =μ j +ε·σ j ,ε~N(0,1),z j Let σ be the j-th parameter of the Z vector. j Let μ be the j-th parameter of the σ vector. j Let ε be the j-th parameter of the μ vector; ε is the parameter that follows a standard normal distribution, with the symbol ~ indicating that it follows a standard normal distribution; N(0,1) is the standard normal distribution; j = 1, 2, ..., d.

[0020] Using Z as the input to the instantaneous prediction time-series memory module and the pose maneuver autonomous control module, the compression feature of vector Z at time t is Zt. t The compression feature at time t+1 is Z. t+1 Z t and Z t+1 As input to step A2, the loss function L of the complex environment autonomous perception module vision As shown in formula (1):

[0021]

[0022] Where x is the input image, For the corresponding reconstruction form, L recon The reconstruction loss is the sum of the input image x and the corresponding reconstructed form. Differences, L KL This is the relative entropy loss function.

[0023] A2. Establish an instantaneous prediction time-series memory module

[0024] Based on the external environment vector obtained in step A1, a hybrid density recurrent neural network is used to perform time-series prediction of the external environment of the fixed-wing UAV. The recurrent neural network outputs in the form of a probability density function P(Z).

[0025] By approximating the probability density function P(Z) as a mixture of Gaussian distributions, a recurrent neural network is trained to output the vector Z at time t+1 using historical and current available information. t+1 The probability distribution is used to realize the function of the instantaneous prediction time-series memory module. The recurrent neural network is modeled as P(Z) t+1 |a t Z t ,h t The operator | indicates that in a t Z t ,h t Vector Z at time t+1 under the given condition t+1 The probability distribution, where a t Z is the behavior of the fixed-wing UAV at time t. t Z is the compressed feature representation of the input image at time t. t ={z1,z2,…,z d} t Each element represents Z t The components of the vector's corresponding dimension at time t, Z t+1 Z is the compressed feature representation of the input image at time t+1. t+1 ={z1,z2,…,z d} t+1 Each element represents Z t+1The components of the dimension of the vector at time t+1, h t h is the latent vector at time t. t ={h1,h2,…,h d} t Each element represents the latent vector h. t The components in the corresponding dimension.

[0026] A3. Establish an autonomous control module for posture maneuvering.

[0027] It accepts the complex environment autonomous perception module and the instantaneous prediction time sequence memory module as input, and maximizes the cumulative reward to generate the corresponding fixed-wing UAV control behavior.

[0028] The inputs to the posture maneuvering autonomous control module include:

[0029] The output Z of the complex environment autonomous perception module t ;

[0030] The latent vector h in the instantaneous prediction time-series memory module t ;

[0031] The output of the attitude maneuvering autonomous control module is a t , representing the behavior of the fixed-wing UAV at time t, as shown in formula (2):

[0032] a t =tanh(ω) c (Z t h t )+b c (2)

[0033] Where tanh(·) is the activation function, ω c and b c These are the weight matrix and the bias vector, respectively. Since the value of tanh(·) ranges from -1 to +1, according to formula (3), a t Normalize:

[0034]

[0035] In the formula, a[0] to a[4] represent the input vector Z. t Mapped to the action vector of the output.

[0036] B. Establish a fast-path adaptive pose control model

[0037] First, a motion mathematical model of the fixed-wing UAV is established to describe its flight state. Then, a pose calculation model of the fixed-wing UAV is established to calculate the pose parameters during its flight. Finally, a stable loop control law for the fixed-wing UAV is designed to obtain the output of the fixed-wing UAV control system.

[0038] The specific steps are as follows:

[0039] B1. Establishing a kinematic mathematical model for fixed-wing unmanned aerial vehicles.

[0040] The flight state of a fixed-wing unmanned aerial vehicle (UAV) is described, and a mathematical model of its motion is established. The attitude and trajectory changes of the fixed-wing UAV are physical processes based on dynamics and kinematics. Different reference frames are used in the mathematical modeling of the fixed-wing UAV, and its navigation parameters, including attitude, position, and velocity, are studied.

[0041] The differential equations of motion of a fixed-wing UAV particle are shown in equation (4):

[0042]

[0043] in, The component of the aerodynamic force on the lift axis. Let be the component of the aerodynamic force on the drag axis. Let ω be the aerodynamic force component on the side force axis, m be the mass of the fixed-wing UAV, ω be the angular velocity vector in the ground inertial coordinate system, v be the velocity vector, g be the gravitational acceleration, F be the thrust generated by the engine, Φ be the roll angle, θ be the pitch angle, u, v, and w be the projection scalars of the velocity vector in the x, y, and z directions of the body coordinate system, and q and r be the projection scalars of the angular velocity vector in the y and z directions of the body coordinate system.

[0044] B2. Establish a fixed-wing UAV attitude calculation model

[0045] The first-order differential equations for the motion of the fixed-wing UAV are shown in equation (5):

[0046]

[0047] Where p, q, and r are the three-axis angular velocities of the body measured by the gyroscope, where p is the x-axis angular velocity, q is the y-axis angular velocity, and r is the z-axis angular velocity; φ, θ, and ψ are Euler angles, where φ is the roll angle, θ is the pitch angle, and ψ is the yaw angle. By solving formula (4), the orthogonal attitude angles of the fixed-wing UAV are obtained, and orthogonalization is no longer required in subsequent navigation calculations.

[0048] The following uses accelerometers and magnetometers for attitude correction. In the absence of self-acceleration in a fixed-wing UAV, the triaxial accelerometer measures the three-axis components of the gravitational field in the body coordinate system to determine the roll and pitch angles of the fixed-wing UAV.

[0049] The expression for the gravity vector in the ground coordinate system e is:

[0050] Ge =[0 0 g]

[0051] When a fixed-wing UAV is not accelerating relative to the ground coordinate system, the readings from the three-axis accelerometer are:

[0052]

[0053] In the formula, a b The acceleration components are in the body coordinate system. The acceleration component in the body coordinate system is the eastern component. The north component of acceleration in the body coordinate system. Let be the vertical component of acceleration in the body coordinate system, and we have:

[0054]

[0055]

[0056] The triaxial magnetometer is installed along the body coordinate system, and the measured values ​​are:

[0057]

[0058] In the ground coordinate system, the geomagnetic intensity value is denoted as:

[0059]

[0060] The formula for calculating the yaw angle is shown in (6):

[0061]

[0062] in, The eastern component of the geomagnetic intensity in the body coordinate system. The northern component of the geomagnetic intensity in the body coordinate system. This represents the vertical component of the geomagnetic intensity in the body coordinate system. This represents the eastern component of the geomagnetic intensity in the ground coordinate system. The northern component of the geomagnetic intensity in the ground coordinate system. The vertical component of geomagnetic intensity is represented in the body coordinate system. The superscript b indicates the body coordinate system, e indicates the ground coordinate system, and the subscripts x, y, and z indicate the east, north, and vertical directions of the geomagnetic field.

[0063] B3. Design of the stability loop control law for fixed-wing unmanned aerial vehicles.

[0064] The flight control system controls the flight status and trajectory of the fixed-wing UAV by controlling the flight parameters, including roll angle, pitch angle, yaw angle, altitude, and latitude and longitude, i.e., the aircraft's position coordinates; among them, altitude and latitude and longitude are used for the trajectory control of the fixed-wing UAV; roll angle, pitch angle, and yaw angle are used for attitude control.

[0065] To improve the performance of the flight control system of a fixed-wing UAV, a nonlinear gain PID controller is adopted. Under small disturbances, the cross-linking effect between the longitudinal and lateral motions of the fixed-wing UAV is not severe. Therefore, the flight control is decomposed into independent longitudinal and lateral control channels. The longitudinal control channel is used to stabilize the pitch angle, altitude, and speed of the fixed-wing UAV; the lateral control channel is used to stabilize the heading angle, roll angle, and yaw distance of the fixed-wing UAV. The output of the control system is shown in formula (7):

[0066]

[0067] Where r(t) is the control command, y(t) is the output quantity, and u(t) is the output of the PID controller.

[0068] In the formula, k p k is the proportional gain parameter. i k is the integral gain parameter. d The differential gain parameter is calculated using a nonlinear function, as shown in formula (8):

[0069]

[0070] C. Establish a brain-like dual-pathway hybrid-driven maneuvering optimization model

[0071] The Epsilon greedy strategy is used to control the motion system of a fixed-wing UAV to optimize maneuvers. The specific steps are as follows:

[0072] C1. Optimize actions using the Epsilon greedy strategy.

[0073] The learning objective of the Epsilon greedy policy is to obtain a policy π during learning, denoted as:

[0074] π:S→A

[0075] Here, S represents a finite set of discrete states, A represents a discrete set of actions, and the symbol → indicates the process of choosing to take action A in state S. At time k in state s... k Below, the selected action is a. n ∈A k The probability distribution is as follows:

[0076] π k ={p1,p2,…,p n}

[0077] Assuming in state s k At that time, the actions adopted by the fixed-wing UAV are derived from the set of behaviors A of the data-driven slow-path instantaneous maneuver obstacle avoidance model.k ={a1,a2,…,a m} and the PID behavioral control quantity U from the model-driven fast-path adaptive pose control model k ={u1,u2,…,u n In the case of hybrid-driven action optimization decision-making, action a is selected at this time. j ,、u j The probability is calculated as shown in formula (9):

[0078]

[0079] Where m is the dimension of the behavior set, n is the dimension of the behavior control quantity, k is the current time, i is the current behavior index, j is the current behavior control quantity index, and Q(s) k ,a i ,u j ) for state-policy group s k a i u j The function is a value function, where ε(k) is a gradual parameter. Its value represents the degree of randomness in action selection; a larger ε(k) indicates greater randomness in the action. M represents the number of learning iterations. When k = 1, ε(k) ≈ 1, representing initial exploration without utilization; when k = M, ε(k) ≈ 0, representing utilization without exploration; 0.99 prevents ε(k) from taking the boundary value 0 or 1, which is an meaningless state. The gradual change of the function ε(k) with k simulates the transition from exploration to utilizing experiential knowledge, i.e., the gradual optimization of the fixed-wing UAV's actions.

[0080] C2. Perform energy-based action jumps.

[0081] When the Epsilon greedy policy performs random action policy transitions, the value function Q(s) for action policy selection... k ,a i ,u j If Q(s) k ,a i+1 ,u j+1 )≥Q(s k ,a i ,u j If a better solution is found after a random jump, then the move is always accepted; if Q(s) k ,a i+1 ,u j+1 )<Q(s k ,a i ,u jIf a random jump results in a worse solution than the current one, then a probability is defined to express the validity of this jump. The probability here is time-dependent, and the probability gradually decreases as the time series progresses, as shown in formula (10). When the jump energy value reaches T, the energy difference is dQ, and the probability is p(dQ).

[0082]

[0083] Where C is a constant, it is known that as the jump energy T gradually decreases, p(dQ(s) k ,a i ,u j The value will also decrease accordingly. Here, each jump to other solution domains is considered an energy jump, denoted by the probability p(dQ(s)). k ,a i ,u j )) indicates the degree to which this jump is accepted.

[0084] Compared with the prior art, the present invention has the following beneficial effects:

[0085] 1. This invention addresses the autonomous obstacle avoidance requirements of fixed-wing UAVs in complex urban and mountainous environments, adapting to highly dynamic environments with low altitudes, high elevation changes, and variable wind fields. It proposes a brain-like fast and slow dual-path autonomous meta-reinforcement learning model-data hybrid-driven dynamic control method system for fixed-wing UAVs. This system utilizes a large amount of offline data generated from repetitive tasks of fixed-wing UAVs and draws on the training and learning mechanisms of maneuvering and obstacle avoidance by manned aircraft pilots. It overcomes the shortcomings of traditional control schemes in terms of intelligent requirements and fully considers the problems faced in maneuvering flight, such as nonlinearity, strong coupling control, Euler angle singularity, and complex maneuver generation, to meet the requirements of high-efficiency and high-precision real-time tracking flight.

[0086] 2. This invention compares the advantages and disadvantages of existing data-driven and model-driven methods, and adopts a brain-like highly autonomous model-data hybrid-driven approach. It consists of three parts: a slow-path data-driven instantaneous maneuvering obstacle avoidance model, a fast-path model-driven adaptive pose control model, and a brain-like fast and slow dual-path model-data hybrid-driven maneuvering motion optimization model. This approach enables stable tracking control of the position and attitude of a fixed-wing UAV and autonomous and coherent completion of maneuvering obstacle avoidance flight.

[0087] 3. This invention adopts a brain-like fast and slow dual-pathway architecture, rationally designs long-term and short-term decision-making game strategies, realizes effective reinforcement mechanisms for each of the fast and slow pathways, implements dynamic game between pathways, and in the final decision-making selection, the two pathways are both relatively independent and mutually reinforcing. This enables the entire brain-like fast and slow dual-pathway model to effectively form intuitive habits through iteration and feedback, thereby improving autonomous cognitive performance and effectively mapping between artificial intelligence models and human brain cognitive mechanisms. It truly heuristically constructs the required autonomous intelligent system based on the interaction mechanism and principle of the fast and slow dual-pathways in the brain and cognition, enabling the autonomous intelligent system to achieve intelligent cognition. Attached Figure Description

[0088] This invention has a total of appendices Figure 4 Zhang, of which:

[0089] Figure 1 This is a flowchart of the present invention.

[0090] Figure 2 This is a diagram of the slow pathway model.

[0091] Figure 3 This is a diagram of a fast-path proportional-derivative-integral control system.

[0092] Figure 4 It is a complex and integrated action diagram. Detailed Implementation

[0093] The invention will now be further described with reference to the accompanying drawings. According to... Figure 1 The flowchart shown introduces the brain-like fast and slow dual-pathway autonomous cognitive obstacle avoidance method. First, as... Figure 2 The following steps are performed sequentially: constructing a complex environment autonomous perception module, an instantaneous prediction and temporal memory module, and a pose maneuvering autonomous control module to establish a slow-path instantaneous maneuvering obstacle avoidance model. Then, a fixed-wing UAV motion mathematical model is constructed to calculate the pose, as shown below. Figure 3 The design of the stabilization loop control law for the fixed-wing UAV is shown, and a fast-path adaptive pose control model is established. Then, a brain-like fast-slow dual-path hybrid drive maneuver optimization model is established, such as... Figure 4 As shown, this enables autonomous obstacle avoidance for fixed-wing UAVs.

[0094] This invention is not limited to this embodiment. Any equivalent concept or modification within the technical scope disclosed in this invention shall be included within the protection scope of this invention.

Claims

1. A model-data hybrid-driven autonomous cognitive obstacle avoidance method for fixed-wing unmanned aerial vehicles, characterized in that: Includes the following steps: A. Establish a slow-path instantaneous maneuvering obstacle avoidance model First, a complex environment autonomous perception module is established to obtain the vector of the fixed-wing UAV to the external environment; then, an instantaneous prediction and timing memory module is established to predict the state of the fixed-wing UAV in a timing sequence; finally, an attitude maneuver autonomous control module is established to generate the optimal control behavior of the fixed-wing UAV; the specific steps are as follows: A1. Establish an autonomous perception module for complex environments. The system observes the external environment using radar and cameras to obtain input images. These images are then compressed frame by frame, and the video frames are encoded into low-dimensional latent vectors using a variational self-encoding and decoding structure, enabling autonomous perception of complex environments. First, the resolution of the input image is compressed to 64×64×3. Then, the image encoding is reduced to a two-dimensional space to obtain μ and σ vectors. The forms of μ and σ vectors are defined as follows: μ = {μ1, μ2, ..., μ d } σ={σ1,σ2,…,σ d } Where d represents the dimensions of the μ and σ vectors; Let f θ f represents the encoding network for the μ vector. φ The formulas for calculating the μ vector and σ vector in the encoding network representing the σ vector are: μ=f θ (x) σ=f φ (x) In the formula, x is the input image; the compression feature Z of the input image is: Z={z1,z2,...,z d } Where: z j =μ j +ε·σ j , ε~N(0,1), z j Let σ be the j-th parameter of the Z vector. j Let μ be the j-th parameter of the σ vector. j Let be the j-th parameter of the μ vector; ε is a parameter that follows a standard normal distribution, with the symbol ~ indicating that it follows a standard normal distribution; N(0,1) is a standard normal distribution; j = 1, 2, ..., d; Using Z as the input to the instantaneous prediction time-series memory module and the pose maneuver autonomous control module, the compression feature of vector Z at time t is Zt. t The compression feature at time t+1 is Z. t+1 Z t and Z t+1 As input to step A2, the loss function L of the complex environment autonomous perception module vision As shown in formula (1): Where x is the input image, For the corresponding reconstruction form, L recon The reconstruction loss is the sum of the input image x and the corresponding reconstructed form. Differences, L KL The relative entropy loss function; A2. Establish an instantaneous prediction time-series memory module Based on the external environment vector obtained in step A1, a hybrid density recurrent neural network is used to perform time-series prediction of the external environment of the fixed-wing UAV; the recurrent neural network outputs in the form of a probability density function P(Z); By approximating the probability density function P(Z) as a mixture of Gaussian distributions, a recurrent neural network is trained to output the vector Z at time t+1 using historical and current available information. t+1 The probability distribution is used to realize the function of the instantaneous prediction time-series memory module; the recurrent neural network is modeled as P(Z) t+1 |a t Z t h t The operator | indicates that in a t Z t h t Vector Z at time t+1 under the given condition t+1 The probability distribution of a, where a t Z is the behavior of the fixed-wing UAV at time t. t Z is the compressed feature representation of the input image at time t. t ={z1, z2, ..., z d } t Each element represents Z t The components of the dimension of the vector at time t, Z t+1 Z is the compressed feature representation of the input image at time t+1. t+1 ={z1, z2, ..., z d } t+1 Each element represents Z t+1 The components of the dimension of the vector at time t+1, h t Let ht be the latent vector at time t, where ht = {h1, h2, ..., ht}. d } t Each element represents the latent vector h. t Components in the corresponding dimension; A3. Establish an autonomous control module for posture maneuvering. It accepts the complex environment autonomous perception module and the instantaneous prediction time sequence memory module as input, and maximizes the cumulative reward to generate the corresponding fixed-wing UAV control behavior; The inputs to the posture maneuvering autonomous control module include: The output Z of the complex environment autonomous perception module t ; The latent vector h in the instantaneous prediction time-series memory module t ; The output of the attitude maneuvering autonomous control module is a t , representing the behavior of the fixed-wing UAV at time t, as shown in formula (2): a t =tanh(ω c (Z t ;h t )+b c ) (2) Where tanh(·) is the activation function, ω c and b c These are the weight matrix and the bias vector, respectively. Since the value of tanh(·) ranges from -1 to +1, according to formula (3), a t Normalize: In the formula, a[0] to a[4] represent the input vector Z. t Mapped to the output action vector; B. Establish a fast-path adaptive pose control model First, a mathematical model of motion for the fixed-wing UAV is established to describe its flight state. Then, a pose calculation model for the fixed-wing UAV is established to calculate its pose parameters during flight. Finally, a stable loop control law for the fixed-wing UAV is designed to obtain the output of the UAV control system. The specific steps are as follows: B1. Establishing a kinematic mathematical model for fixed-wing unmanned aerial vehicles. The flight state of a fixed-wing UAV is described, and a mathematical model of its motion is established. The attitude and trajectory changes of the fixed-wing UAV are physical processes based on dynamics and kinematics. Different reference frames are used when mathematically modeling the fixed-wing UAV. The navigation parameters of the fixed-wing UAV are studied, including attitude, position and velocity. The differential equations of motion of a fixed-wing UAV particle are shown in equation (4): in, The component of the aerodynamic force on the lift axis. Let be the component of the aerodynamic force on the drag axis. Let ω be the aerodynamic force component on the side force axis, m be the mass of the fixed-wing UAV, ω be the angular velocity vector in the ground inertial coordinate system, v be the velocity vector, g be the gravitational acceleration, F be the thrust generated by the engine, Φ be the roll angle, θ be the pitch angle, u, v, and w be the projection scalars of the velocity vector in the x, y, and z directions of the body coordinate system, and q and r be the projection scalars of the angular velocity vector in the y and z directions of the body coordinate system, respectively. B2. Establish a fixed-wing UAV attitude calculation model The first-order differential equations for the motion of the fixed-wing UAV are shown in equation (5): Where p, q, and r are the three-axis angular velocities of the body measured by the gyroscope, where p is the x-axis angular velocity, q is the y-axis angular velocity, and r is the z-axis angular velocity; φ, θ, and ψ are Euler angles, where φ is the roll angle, θ is the pitch angle, and ψ is the yaw angle; by solving formula (4), the orthogonal attitude angles of the fixed-wing UAV are obtained, and orthogonalization is no longer required in subsequent navigation calculations; The following uses accelerometers and magnetometers for attitude correction; when the fixed-wing UAV does not have its own acceleration, the three-axis accelerometer determines the roll and pitch angles of the fixed-wing UAV by measuring the three-axis components of the gravitational field in the body coordinate system. The expression for the gravity vector in the ground coordinate system e is: G e =[0 0 g] When a fixed-wing UAV is not accelerating relative to the ground coordinate system, the readings from the three-axis accelerometer are: In the formula, a b The acceleration components are in the body coordinate system. The acceleration component in the body coordinate system is the eastern component. The north component of acceleration in the body coordinate system. Let be the vertical component of acceleration in the body coordinate system, and we have: The triaxial magnetometer is installed along the body coordinate system, and the measured values ​​are: In the ground coordinate system, the geomagnetic intensity value is denoted as: The formula for calculating the yaw angle is shown in (6): in, The eastern component of the geomagnetic intensity in the body coordinate system. The northern component of the geomagnetic intensity in the body coordinate system. This represents the vertical component of the geomagnetic intensity in the body coordinate system. This represents the eastern component of the geomagnetic intensity in the ground coordinate system. The northern component of the geomagnetic intensity in the ground coordinate system. The vertical component of geomagnetic intensity is represented in the body coordinate system. The superscript b indicates the body coordinate system, e indicates the ground coordinate system, and the subscripts x, y, and z indicate the east, north, and vertical directions of the geomagnetic field. B3. Design of the stability loop control law for fixed-wing unmanned aerial vehicles. The flight control system controls the flight status and trajectory of the fixed-wing UAV by controlling the flight parameters, including roll angle, pitch angle, yaw angle, altitude, and latitude and longitude, i.e., the aircraft's position coordinates; among them, altitude and latitude and longitude are used for the fixed-wing UAV's trajectory control; roll angle, pitch angle, and yaw angle are used for attitude control. To improve the performance of the flight control system of a fixed-wing UAV, a nonlinear gain PID controller is adopted. Under small disturbances, the cross-linking effect between the longitudinal and lateral motions of the fixed-wing UAV is not serious. Therefore, the flight control is decomposed into independent longitudinal control channels and lateral control channels. The longitudinal control channel is used to stabilize the pitch angle, altitude, and speed of the fixed-wing UAV. The lateral control channel is used to stabilize the heading angle, roll angle, and yaw distance of the fixed-wing UAV. The output of the control system is shown in formula (7). Where r(t) is the control command, y(t) is the output quantity, and u(t) is the output of the PID controller; In the formula, k p k is the proportional gain parameter. i k is the integral gain parameter. d The differential gain parameter is calculated using a nonlinear function, as shown in formula (8): C. Establish a brain-like dual-pathway hybrid-driven maneuvering optimization model The Epsilon greedy strategy is used to control the motion system of a fixed-wing UAV to optimize maneuvers. The specific steps are as follows: C1. Optimize actions using the Epsilon greedy strategy. The learning objective of the Epsilon greedy policy is to obtain a policy π during learning, denoted as: π: S→A Here, S represents a finite set of discrete states, A represents a discrete set of actions, and the symbol → indicates the process of choosing to take action A in state S; at time k in state s k Below, the selected action is a. n ∈A k The probability distribution is as follows: p k ={p1,p2,…,p n }; Assuming in state s k At that time, the actions adopted by the fixed-wing UAV are derived from the set of behaviors A of the data-driven slow-path instantaneous maneuver obstacle avoidance model. k ={a1, a2, ..., a m } and the PID behavioral control quantity U from the model-driven fast-path adaptive pose control model k ={u1, u2, ..., u n In the case of hybrid-driven action optimization decision-making, action a is selected at this time. j , u j The probability is calculated as shown in formula (9): Where m is the dimension of the behavior set, n is the dimension of the behavior control quantity, k is the current time, i is the current behavior index, j is the current behavior control quantity index, and Q(s) k a i u j ) for state-policy group s k a i u j The value function is defined by the parameter ε(k), which is a gradual parameter. Its value represents the degree of randomness in action selection. The larger the parameter ε(k), the greater the randomness of the action. M is the number of learning iterations. When k = 1, ε(k) ≈ 1 represents that only exploration is performed at the beginning of learning without utilization. When k = M, ε(k) ≈ 0 represents that only utilization is performed without exploration. 0.99 prevents ε(k) from taking the boundary value of 0 or 1, which is a meaningless state. The gradual change of the function ε(k) with k simulates the transition from exploration to utilization of experience knowledge, that is, the process of gradual optimization of the fixed-wing UAV's actions. C2. Perform energy-based action jumps. When the Epsilon greedy policy performs random action policy transitions, the value function Q(s) for action policy selection... k a i u j If Q(s) k a i+1 u j+1 )≥Q(s k a i u j If a better solution is found after a random jump, then the jump is always accepted; if Q(s) k a i+1 u j+1 )<Q(s k a i u j If a solution is obtained after a random jump, then a probability is defined to express the validity of this jump. The probability here is related to time, and the probability gradually decreases as the time series progresses, as shown in formula (10). When the jump energy value reaches T, the energy difference is dQ and the probability is p(dQ). Where C is a constant, it is known that as the jump energy T gradually decreases, p(dQ(s) k a i u j The value will also decrease accordingly; here, each jump to other solution domains is considered an energy jump, using the probability p(dQ(s)). k a i u j )) indicates the degree to which this jump is accepted.

Images