Multi-unmanned aerial vehicle cooperative crossing method and device

Abstract

The invention belongs to the technical field of unmanned aerial vehicle cooperative crossing, and relates to a multi-unmanned aerial vehicle cooperative crossing method and device. The method comprises the following steps: acquiring scene information of cooperative crossing of multiple unmanned aerial vehicles, and obtaining a data set, a model of a crossing frame and a dynamic model of each unmanned aerial vehicle; according to the dynamic model of each unmanned aerial vehicle, constructing a model prediction control framework of each unmanned aerial vehicle; acquiring a historical trajectory sequence of each unmanned aerial vehicle to obtain a trained neural network, and embedding the trained neural network into a model prediction control framework to form a distributed model prediction control framework of multiple unmanned aerial vehicles; and performing path prediction planning on the current unmanned aerial vehicle and other unmanned aerial vehicles according to the distributed model prediction control framework of the current unmanned aerial vehicle, the model of the crossing frame and the kinetic model of the current unmanned aerial vehicle to obtain the real-time position of each unmanned aerial vehicle so as to perform multi-unmanned aerial vehicle cooperative crossing. According to the invention, cooperative crossing can be realized under the condition that multiple unmanned aerial vehicles do not communicate with each other.

Description

Technical Field

[0001] The present application relates to the technical field of cooperative crossing of unmanned aerial vehicles (UAVs), and in particular to a method and device for cooperative crossing of multiple UAVs. Background Art

[0002] With the development of science and technology and the advancement of technology, the coordinated flight of multiple UAV systems has been studied rapidly.

[0003] In existing technologies, ensuring that multiple drones can traverse in an agile, precise, and orderly manner when performing collaborative flight missions is a challenging research problem. This is especially true in scenarios with dynamic or static door frames. Multiple drones not only need to efficiently plan the optimal path, but also need to avoid collisions with each other and complete the target mission within a limited time.

[0004] However, this problem involves multiple core research directions such as drone dynamics modeling, motion planning, path optimization, obstacle avoidance strategies, etc., and requires the combination of real-time optimization algorithms and predictive control methods to achieve robust and efficient trajectory planning. At present, relevant technologies have not yet appeared. Summary of the Invention

[0005] Based on this, it is necessary to provide a multi-UAV collaborative crossing method and device to address the above technical problems, which can enable multiple UAVs to achieve collaborative crossing without communicating with each other.

[0006] A multi-UAV collaborative traversal method, comprising: Obtain scene information of multi-UAV collaborative traversal, collect information, and generate data sets; According to the scene information of multi-UAV collaborative crossing, the crossing frame is modeled to obtain the crossing frame model; Based on the scene information of multi-UAV collaborative crossing, each UAV is modeled to obtain the dynamic model of each UAV; Based on the dynamic model of each UAV and combined with the quadratic programming optimization method, a model predictive control framework for each UAV is constructed; The historical trajectory sequence of each drone is obtained, and a neural network is constructed. The neural network is trained using the dataset as input to obtain a trained neural network. The trained neural network is embedded in the model predictive control framework to form a distributed model predictive control framework for multiple drones. Based on the distributed model predictive control framework of the current drone, the model of the traversal box, and the dynamic model of the current drone, path prediction and planning are performed for the current drone and other drones, the current drone's actions are output, and all drones are traversed to obtain the real-time position of each drone. Based on the real-time position of each drone, multi-drone collaborative crossing is carried out.

[0007] In one embodiment, obtaining scene information of multiple UAVs cooperating to traverse, collecting information, and generating a data set include: Obtain scene information of multi-UAV collaborative traversal and obtain the initial environment state; Initial observation information and initial state information are obtained based on the initial environment state; a corresponding Gaussian strategy is constructed for each drone based on the initial observation information and initial state information; random sampling is performed in each drone's Gaussian strategy to obtain each drone's initial action and form a set of candidate actions; the candidate actions interact with the environment to obtain reward indicators and pass indicators; based on the reward indicators and pass indicators, the sample weights are calculated and the Gaussian strategy of each drone is updated; new candidate actions are formed based on the updated strategy, and the new candidate actions, initial environment state, initial observation information, and initial state information are used as data at the current sampling time; Update the environment state and send the new candidate action to the environment to drive the system update, obtain the next environment state, the next observation information, and the next state information; obtain the next candidate action based on the next observation information and the next state information, and use the next candidate action, the next environment state, the next observation information, and the next state information as the data for the next sampling time; The environmental status is updated cyclically until the end of the sampling period, and a data set is generated based on the data at all sampling moments.

[0008] In one embodiment, a traversal frame is modeled based on the scene information of the multi-UAV collaborative traversal, and the traversal frame model is obtained, including: Based on the scene information of multi-UAV collaborative crossing, the state vector of the crossing box is defined; Perform force analysis on the crossing frame and obtain the change in the swing angle of the crossing frame based on the state vector of the crossing frame; According to the change of the swing angle of the crossing frame, the expression of the center position of the crossing frame is obtained; Based on the expression of the center position of the crossing frame, the expression of the corner point position of the crossing frame is obtained; According to the expression of the corner point position of the crossing frame, the center position of the crossing frame is obtained and used as the model of the crossing frame.

[0009] In one embodiment, based on the scene information of the multi-UAV cooperative traversal, each UAV is modeled to obtain a dynamic model of each UAV, including: According to the scene information of multi-UAV collaborative traversal, the state vector of each UAV is obtained; According to the state vector of each UAV, under the constraint of the objective function, the control vector of each UAV is obtained; According to the control vector of each UAV, the linear velocity change rate and attitude change rate of each UAV are obtained; According to the linear velocity change rate and attitude change rate of each UAV, dynamic modeling is performed on each UAV to obtain the dynamic model of each UAV.

[0010] In one embodiment, the objective function includes: ; in, ; ; ; ; Where, is the objective function, is the target approach term, is the trajectory tracking term, To control the smoothing term, is the collaborative collision avoidance term, For the The state of the moment, is the desired end state, is a positive definite matrix, For the time, To predict the total number of steps, is the time weighting factor, For the The first drone The predicted state of the step, For the The reference state of the step, is the state error weight matrix, For the The drone The control input of the step, is the control input in the hover state, is the control weight matrix, is the predicted number of steps of other drones, For the drones, For the drones, is the collision avoidance gain parameter, is the current position of the drone, For the The drone The position of the step, Is a positive number.

[0011] In one embodiment, a model predictive control framework for each UAV is constructed based on the dynamic model of each UAV and combined with a quadratic programming optimization method, including: ; Where, is the dynamic model of the UAV, is the state vector of the UAV, is the position coordinate of the UAV in the inertial coordinate system, is the attitude quaternion corresponding to the drone, is the linear velocity component of the UAV in the inertial coordinate system, is the control vector of the UAV; is the total thrust in the body coordinate system, is the angular velocity component in the body coordinate system; Perform Euler discretization: ; Where, For drones The state of the moment, For drones The state of the moment, To control the time step, For control input.

[0012] In one embodiment, the historical trajectory sequence of each drone is obtained and a neural network is constructed, including: Obtain the historical trajectory sequence and map it to a high-dimensional feature space through a linear embedding layer to obtain trajectory data; Design an encoder and stack it with a multi-head self-attention mechanism and a feedforward network to model the spatiotemporal dependencies in the trajectory data and obtain the encoder output; The output of the encoder is used as the input of the residual enhanced long short-term memory network for temporal modeling; Design a multi-scale causal convolution module to process the results of time series modeling to adjust the drone's fine-tuning within a short time window and obtain the initial neural network; Construct a multi-objective loss function of the initial neural network and obtain the neural network.

[0013] In one embodiment, the neural network is trained using the data set as input to obtain a trained neural network, including: ; Where, To predict the future The output of the step, is the feature fusion function, is the feature extraction function of Transformer, is the feature extraction function of the long short-term memory network, is the feature extraction function of CNN, From time 1 to The input sequence of .

[0014] In one embodiment, the trained neural network is embedded into the model predictive control framework to form a distributed model predictive control framework for multiple drones, including: The trained neural network is embedded into the model predictive control framework to add the status and quantity of other drones to the model predictive control framework of the current drone, thus obtaining a distributed model predictive control framework for the current drone. Based on the distributed model predictive control framework of all drones, a distributed model predictive control framework for multiple drones is formed.

[0015] A multi-UAV cooperative crossing device, comprising: The first module is used to obtain scene information of multi-UAV collaborative traversal, collect information, and generate a data set; The second module is used to model the crossing frame based on the scene information of the multi-UAV collaborative crossing, and obtain the crossing frame model; The third module is used to model each UAV based on the scene information of multi-UAV collaborative crossing and obtain the dynamic model of each UAV; The fourth module is used to build a model predictive control framework for each UAV based on its dynamic model and combined with quadratic programming optimization method; The fifth module is used to obtain the historical trajectory sequence of each drone, construct a neural network, and use the dataset as the input to train the neural network to obtain a trained neural network. The trained neural network is embedded in the model predictive control framework to form a distributed model predictive control framework for multiple drones. Based on the distributed model predictive control framework of the current drone, the model of the traversal box, and the dynamic model of the current drone, the path prediction planning is performed for the current drone and other drones, the action of the current drone is output, and all drones are traversed to obtain the real-time position of each drone. The sixth module is used to conduct multi-UAV collaborative crossing based on the real-time position of each UAV.

[0016] The above-mentioned multi-UAV collaborative crossing method and device is a trajectory optimization method based on model predictive control (MPC). Through spatiotemporal decoupling modeling (multi-head attention), dynamic gated fusion (Transformer-LSTM-CNN) and physical constraint injection, it achieves high precision and strong robustness in UAV trajectory prediction, while meeting the dynamic constraint requirements of the actual flight control system, providing a reliable state estimation basis for the autonomous navigation system, enabling multiple UAVs to complete the crossing task safely and efficiently under the constraints of dynamic and static obstacles. It is particularly suitable for the collaborative crossing scenario of multiple UAVs under the constraints of dynamic and static obstacles. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 Schematic diagram of a process of a multi-UAV collaborative traversal method according to an embodiment; Figure 2 A network structure diagram of a multi-UAV cooperative traversal method in one embodiment; Figure 3 A schematic diagram of a scenario of a multi-UAV cooperative traversal method according to an embodiment; Figure 4 The figure is a structural block diagram of a multi-UAV cooperative crossing device in one embodiment. DETAILED DESCRIPTION

[0018] In order to make the purpose, technical solutions and advantages of this application more clearly understood, the present application is further described in detail below with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are merely for explaining this application and are not intended to limit this application. All other embodiments obtained by persons of ordinary skill in the art based on the embodiments in this application without creative work are within the scope of protection of this application.

[0019] In addition, the terms "first," "second," and so on, used in this application are for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Therefore, features specified as "first" or "second" may explicitly or implicitly include at least one such feature. In the description of this application, "multiple groups" means at least two groups, such as two groups, three groups, and so on, unless otherwise specifically defined.

[0020] In this application, unless otherwise specified or limited, the terms "connect," "fix," etc. should be understood in a broad sense. For example, "fix" can mean a fixed connection, a detachable connection, or an integral connection; it can mean a mechanical connection, an electrical connection, a physical connection, or a wireless communication connection; it can mean a direct connection or an indirect connection through an intermediate medium; it can mean internal communication between two elements or an interaction between two elements, unless otherwise specified. Those skilled in the art will understand the specific meanings of the above terms in this application based on the specific circumstances.

[0021] In addition, the technical solutions between the various embodiments of the present application can be combined with each other, but it must be based on the fact that ordinary technicians in this field can implement it. When the combination of technical solutions is mutually contradictory or cannot be implemented, it should be deemed that such a combination of technical solutions does not exist and is not within the scope of protection required by this application.

[0022] This application provides a multi-UAV collaborative crossing method, such as Figure 1 The flowchart shown, in one embodiment, includes: Step 101: Acquire scene information of multi-UAV collaborative traversal, collect information, and generate a data set.

[0023] Specifically: Obtain scene information of multi-UAV collaborative traversal and obtain the initial environment state; Initial observation information and initial state information are obtained based on the initial environment state; a corresponding Gaussian strategy is constructed for each drone based on the initial observation information and initial state information; random sampling is performed in each drone's Gaussian strategy to obtain each drone's initial action and form a set of candidate actions; the candidate actions interact with the environment to obtain reward indicators and pass indicators; based on the reward indicators and pass indicators, the sample weights are calculated and the Gaussian strategy of each drone is updated; new candidate actions are formed based on the updated strategy, and the new candidate actions, initial environment state, initial observation information, and initial state information are used as data at the current sampling time; Update the environment state and send the new candidate action to the environment to drive the system update, obtain the next environment state, the next observation information, and the next state information; obtain the next candidate action based on the next observation information and the next state information, and use the next candidate action, the next environment state, the next observation information, and the next state information as the data for the next sampling time; The environmental status is updated cyclically until the end of the sampling period, and a data set is generated based on the data at all sampling moments.

[0024] More specifically: Obtain the scene information of the multi-UAV collaborative traversal, call the environment reset function (this function is the existing technology), and obtain the initial environment state (which can also be considered as the state of the traversal frame); Based on the initial environmental state, the states of all drones are initialized to obtain initial observation information and initial state information. The initial observation information is the 10-dimensional state vector of each drone, including position, attitude (represented by quaternions), and linear velocity. The initial state information is the 4-dimensional motion vector of each drone, including thrust and three angular velocities. In addition, by obtaining the upper and lower bounds of the motion space (i.e., the preset motion boundaries), the generated motion can be clipped to ensure that the generated motion is within the range allowed by physics and actuators. Based on the initial observation information and initial state information, a corresponding Gaussian strategy is constructed for each UAV, i.e., strategy initialization. The action distribution of this strategy takes the initial observation information and initial state information at the current moment as input, and is clipped and generated by combining the lower and upper bounds of the action space. This is used to subsequently generate candidate actions for subsequent optimization. For each drone, within the specified maximum number of iterations (max_wml_iter), randomly sample from the Gaussian policy to obtain each drone's initial action and form a set of candidate actions; interact multiple sets (n_samples) of candidate actions with the environment and execute these actions in the environment to obtain corresponding reward feedback, obtain reward indicators and pass indicators, and use the reward results to evaluate each action sample to subsequently select the optimal action and policy update; For each drone, a sample weight (i.e., a benchmark factor) is calculated based on the candidate actions and their corresponding reward indicators. The sample weights are normalized using rewards, giving higher weights to samples with higher rewards. The mean and covariance matrices of the strategy are updated using the Reward-Weighted Maximum Likelihood method, making the new strategy more likely to generate high-reward actions during sampling. If the reward value of the updated strategy increases, the passing indicator is met, and the strategy update is considered successful. Otherwise, the strategy is not updated, and the previous optimal strategy or heuristic backup solution is still used as the current strategy. According to the updated strategy, sampling is performed to form new candidate actions as the optimal action at the current moment. The new candidate actions (including the actions of each drone), the initial environment state (the state of the traversal box), the initial observation information (the drone's state vector), and the initial state information (the drone's action vector) are used as the data at the current sampling moment. Update the environment state and send the new candidate action to the environment to drive the system update, obtain the next environment state, the next observation information, and the next state information; obtain the next candidate action based on the next observation information and the next state information, and use the next candidate action, the next environment state, the next observation information, and the next state information as the data for the next sampling time; Perform environmental stepping (i.e., call the environment's stepping function, specifically the model of the crossing box in step 102) to cyclically update the environmental state (i.e., update the state of the crossing box center point according to the crossing box model) until the end of the sampling period. Count the number of samples to determine whether the sampling limit has been reached, and generate a data set based on the data at all sampling moments. .

[0025] In this step, the collected dataset will be stored in a dedicated dataset object for subsequent offline training and evaluation.

[0026] In this step, a data collection method based on environmental interaction combined with weighted maximum likelihood is designed, see Table 1:

[0027] As shown in Table 1, through multiple sampling and online policy updates, a training dataset is efficiently constructed to support policy learning for multi-UAV collaborative control. This enables comprehensive collection of UAV states, actions, and corresponding feedback rewards in a simulation environment, providing data support for subsequent policy learning and performance evaluation.

[0028] Step 102: Model the traversal frame based on the scene information of the multi-UAV collaborative traversal to obtain a traversal frame model.

[0029] Specifically: Based on the scene information of multi-UAV collaborative crossing, the state vector of the crossing box is defined; Perform force analysis on the crossing frame and obtain the change in the swing angle of the crossing frame based on the state vector of the crossing frame; According to the change of the swing angle of the crossing frame, the expression of the center position of the crossing frame is obtained; Based on the expression of the center position of the crossing frame, the expression of the corner point position of the crossing frame is obtained; According to the expression of the corner point position of the crossing frame, the center position of the crossing frame is obtained and used as the model of the crossing frame.

[0030] More specifically: To accurately simulate the traversal behavior of a UAV in a dynamic environment, a moving frame structure based on a simple pendulum model was constructed. This structure, called the traversal frame, dynamically changes through the rotation of its pivot point. This effectively simulates the behavior of dynamic obstacles in real-world scenarios, such as a swinging door frame. The following details the modeling of the traversal frame, including its simple pendulum kinematics, geometric structure, and dynamic characteristics. According to the scene information of multi-UAV collaborative crossing, the state vector of the crossing box is defined: ; Where, is the swing angle of the frame (offset relative to the vertical direction), is the angular velocity of the frame; Perform force analysis on the through-frame (because it is only subject to gravity, the motion of the frame is described by the simple pendulum dynamics equation), and obtain the change in the swing angle of the through-frame based on the state vector of the through-frame: ; Where, is the acceleration due to gravity, , is the length of the pendulum, that is, the distance from the suspension point to the center of the frame, is the damping coefficient, which is used to describe the energy loss caused by friction or air resistance during the swing of the frame. is the mass of the frame; According to the change in the swing angle of the crossing frame, the expression for the center position of the crossing frame is obtained: ; ; ; Where, is the fulcrum position; Based on the expression of the center position of the crossing box, the expression of the corner position of the crossing box is obtained: ; ; ; ; Where, is the position of the first corner point, is the position of the second corner point, is the position of the third corner point, is the position of the fourth corner point, is the inner ring width of the crossing frame, is the inner ring height of the crossing frame, For the central location; According to the expression of the corner point position of the crossing frame, the center position of the crossing frame is obtained and used as the model of the crossing frame.

[0031] In this step, the crossing box is modeled.

[0032] Step 103 : Model each UAV based on the scene information of the multi-UAV collaborative crossing to obtain a dynamic model of each UAV.

[0033] Specifically: According to the scene information of multi-UAV collaborative traversal, the state vector of each UAV is obtained; According to the state vector of each UAV, under the constraint of the objective function, the control vector of each UAV is obtained; According to the control vector of each UAV, the linear velocity change rate and attitude change rate of each UAV are obtained; According to the linear velocity change rate and attitude change rate of each UAV, dynamic modeling is performed on each UAV to obtain the dynamic model of each UAV.

[0034] More specifically: To model the UAV, the following basic assumptions are made: 1) Rigid body assumption: The drone is considered as an ideal rigid body, and its deformation and flexible vibration are negligible; 2) Uniform mass distribution: The body is symmetrical and the mass is evenly distributed at the center; 3) Ignore air resistance: The effect of aerodynamic drag on the drone is not considered (aerodynamic effects can be included in subsequent modeling); 4) Motor transient response: The motor dynamics model is not explicitly modeled, and it is assumed that the motor can respond instantaneously to the angular velocity control command.

[0035] To describe the motion of the drone, two coordinate systems are used, one is the inertial coordinate system (ground fixed reference system): , used to describe the absolute position and velocity of the UAV, where In the world coordinate system axis, In the world coordinate system axis, In the world coordinate system Axis; the other is the body coordinate system: ,in, In the body coordinate system Axis, pointing to the drone's nose, In the body coordinate system axis, In the body coordinate system Axis. The transformation between the two coordinate systems is done by the rotation matrix or quaternion express.

[0036] According to the scene information of multi-UAV collaborative traversal, the state vector of each UAV is obtained: ; Where, is the position coordinate of the UAV in the inertial coordinate system, (where, The UAV is in the inertial coordinate system Location coordinates, The UAV is in the inertial coordinate system Location coordinates, The UAV is in the inertial coordinate system Position coordinates), is the attitude quaternion corresponding to the drone (where, The real part of the UAV attitude quaternion is used to represent the body attitude. is the imaginary part of the drone attitude quaternion Quantity, is the imaginary part of the drone attitude quaternion Quantity, is the imaginary part of the drone attitude quaternion Components) (Compared to Euler angles, quaternions can avoid the universal lock problem and have better numerical stability in optimization calculations), satisfying the unit constraint , is the linear velocity component of the UAV in the inertial coordinate system (where, The UAV is in the inertial coordinate system Directional linear velocity component, The UAV is in the inertial coordinate system Directional linear velocity component, The UAV is in the inertial coordinate system Directional linear velocity component); According to the state vector of each UAV, under the constraint of the objective function, the control vector of each UAV is obtained: ; Where, is the total thrust in the body coordinate system (the translational motion of the drone is determined by the external force, among which the total thrust Provided by four rotors, acting on the body axis direction, controlling the lift and vertical acceleration of the drone), is the angular velocity component in the body coordinate system (where In the body coordinate system Directional angular velocity component, In the body coordinate system Directional angular velocity component, In the body coordinate system Directional angular velocity components), the three angular velocity components form the angular velocity , used to adjust the attitude of the drone; According to the control vector of each UAV, the linear velocity change rate and attitude change rate of each UAV are obtained; Among them, the translational motion of the UAV is determined by the external force, where the total thrust Acting on the body coordinate system, gravity Acting on the inertial coordinate system, the acceleration of the drone in the inertial coordinate system is: ; in, ; The linear velocity conversion rate of the UAV is: ; Where, is the rotation matrix, calculated from the quaternion, is the total thrust in the body coordinate system, is gravity, acting on the inertial coordinate system; The attitude change of the drone is determined by the angular velocity Control, its quaternion derivative is: ; Where, is the quaternion derivative; By quaternion derivative Obtaining the attitude change rate of the drone (how to obtain it specifically belongs to the existing technology); According to the linear velocity change rate and attitude change rate of each UAV, the dynamic model of each UAV is constructed to obtain the dynamic model of each UAV: ; Where, The dynamic model of the UAV.

[0037] Among them, in the traditional MPC model, the controller solves the optimal control input based on the state prediction of a single system. For multi-UAV collaborative tasks, in order to achieve decentralization and real-time collaborative control, this application proposes a data-driven distributed model predictive control framework. The framework independently constructs a local optimization problem within each UAV, and at the same time, by introducing the state prediction information of neighboring UAVs, it realizes information coupling and collaborative constraints between multiple UAVs, thereby simultaneously meeting multiple control objectives such as target tracking, input smoothing and collision avoidance. The objective function includes: Under the distributed MPC framework, each UAV independently constructs a local optimization problem, and its decision variables include three parts: 1) its own state trajectory : Describing the future The state evolution at each moment, starting from the current state, satisfies the discrete state transfer equation; 2) Control input sequence : Contains the future The control instructions at the moment must meet the physical and actuator input upper and lower bounds, such as the extreme value limits of thrust and angular velocity; 3) Neighbor state prediction parameters: In order to achieve multi-machine coordination and collision avoidance, each UAV receives additional prediction parameters of other UAVs in a limited number of prediction steps (denoted as These states are rearranged and concatenated and then passed as optimization parameters, so that each local problem can consider the impact of the dynamics of neighboring drones on its own planning; Based on the above variables, the objective function of the local optimization problem for each UAV is established by comprehensively considering the three major indicators of approaching the target, trajectory tracking, control smoothness, and collaborative collision avoidance: ; in, ; ; ; ; Where, is the objective function, For the target approach term, define the terminal cost so that the UAV predicts that the terminal is close to the desired target. For the trajectory tracking item, define the trajectory tracking cost to enable the drone to smoothly and accurately pass through the moving door frame. To control the smoothing term (in order to avoid the execution instability caused by drastic changes in the control command, a quadratic penalty is imposed on the part of the control input that deviates from the preset benchmark, thereby introducing the control smoothing term), the control smoothing cost is defined to suppress the drastic fluctuations of the control input. For the collaborative collision avoidance term (to ensure that UAVs maintain a safe distance during collaborative operation, a collaborative collision avoidance term, namely the collision avoidance cost term, is constructed based on the potential field function theory. When UAVs approach each other, the cost rises rapidly, thereby automatically driving the system to choose a strategy to maintain a safe distance during the optimization process), the collision avoidance cost is defined. By using the state prediction of the adjacent UAVs Achieve safe coordination among multiple machines. For the The state of the moment, is the desired end state, is a positive definite matrix that gives sufficient penalty to the state error, For the time, To predict the total number of steps, is the time weighting factor, ,in, Controls the decay rate, Indicates a critical moment when higher precision control is expected; For the The first drone The predicted state of the step, For the The reference state of the step, is the state error weight matrix, For the The drone The control input of the step, For control input in the hover state (such as ), To control the weight matrix to balance tracking performance and control smoothness, is the predicted number of steps of other drones, , represents the number of steps to consider the influence of neighboring machines in the prediction interval; For the drones, For the drones, is the collision avoidance gain parameter, For the The current location of the drone, For the The drone The position of the step, Be positive in case the denominator is zero.

[0038] In this step, the dynamic model of each UAV is established.

[0039] Step 104 : Based on the dynamic model of each UAV and in combination with the quadratic programming optimization method, a model predictive control framework for each UAV is constructed.

[0040] Specifically: A linearized UAV dynamics model is used in conjunction with a quadratic programming (QP) optimization method to construct a Model Predictive Control (MPC) framework. This framework optimizes control inputs within a finite time domain, guiding the UAV along a predetermined trajectory in an optimal manner. This allows for optimal trajectory planning and real-time trajectory tracking. In the MPC framework, the discrete-time dynamics model is used for prediction, and the continuous-time dynamics equation of the UAV is: ; Where, is the dynamic model of the UAV, is the state vector of the UAV, is the control vector of the UAV (as input); In the MPC framework, it is discretized using Euler: ; in, For drones at all times status, For drones at all times status, To control the time step, For control input.

[0041] In this step, the model predictive control framework for each UAV is constructed.

[0042] Step 105: Obtain the historical trajectory sequence of each UAV, construct a neural network, and use the data set as the input of the neural network to train the neural network to obtain a trained neural network; embed the trained neural network into the model predictive control framework to form a distributed model predictive control framework for multiple UAVs; based on the distributed model predictive control framework of the current UAV, the model of the crossing box, and the dynamic model of the current UAV, perform path prediction planning for the current UAV and other UAVs, output the action of the current UAV, and traverse all UAVs to obtain the real-time position of each UAV.

[0043] Specifically: Acquire historical trajectory sequences and map them to a high-dimensional feature space through a linear embedding layer to obtain trajectory data. Design an encoder and stack it with a multi-head self-attention mechanism and a feedforward network to model the spatiotemporal dependencies in the trajectory data, obtaining the encoder output. Use the encoder output as the input of a residual-enhanced long short-term memory network for time series modeling. Design a multi-scale causal convolution module to process the results of time series modeling to fine-tune the drone's movements within a short time window, obtaining an initial neural network. Construct a multi-objective loss function for the initial neural network to obtain the neural network. Use the data set as the input of the neural network to train the neural network and obtain a trained neural network; The trained neural network is embedded into the model predictive control framework to add the status and quantity of other drones to the model predictive control framework of the current drone, thereby obtaining a distributed model predictive control framework for the current drone. Based on the distributed model predictive control frameworks of all drones, a distributed model predictive control framework for multiple drones is formed. According to the distributed model predictive control framework of the current UAV, the model of the crossing box and the dynamic model of the current UAV, the path prediction planning is performed for the current UAV and other UAVs, the action of the current UAV is output, and all UAVs are traversed to obtain the real-time position of each UAV.

[0044] More specifically: Get the historical trajectory sequence of each drone ,in, is the batch size, is the sequence length, 10 is the dimension, including the position vector , attitude quaternion and velocity vector , and mapped it to a high-dimensional feature space through a linear embedding layer to improve the representation ability of the original state quantity and obtain trajectory data: ; Where, To learn the weight matrix, map features to hidden dimensions , is the bias term; After obtaining the high-dimensional embedding, a Transformer encoder is designed to model the spatiotemporal dependencies in the trajectory data. A multi-head self-attention mechanism (Multi-Head Attention) is stacked with a feedforward neural network (FFN) in the encoder layer to capture the complex spatiotemporal interactions in the drone trajectory. This modeling of the spatiotemporal dependencies in the trajectory data yields the encoder output. The multi-head self-attention mechanism can enhance the ability to model the temporal dependencies between drone states, as shown below: ; Among them, the query matrix , key matrix , value matrix Obtained by linear transformation of the input state; The feedforward network is used to transform features and enhance the nonlinear expression ability of the model, which is expressed as follows: ; Where, is the input of the network structure; represents a nonlinear activation function; are different weight matrices; For different paranoid matrices; The encoder output is used as the input of a residual-enhanced long short-term memory network (residual-enhanced LSTM). This residual-enhanced LSTM is used to process the temporal inertia characteristics of the drone trajectory to perform collaborative modeling of global features and local memory to achieve temporal modeling. The core calculation formula is as follows: ; Where, are the activation values ​​of the forget gate, input gate, and output gate respectively; For candidate memory units, the current context information of Transformer is introduced to enhance the adaptability to complex environments; For the updated memory unit, explicitly record the drone's motion history; are different weight matrices; are different bias terms; is the activation function; tan h represents the hyperbolic tangent function; is a logical operator; For the The hidden state of the moment; is the output of transformer at time t; For the The final output at the moment represents the current state; Because LSTM has a response lag when dealing with local mutations and short-term dynamic changes, it is difficult to accurately model small disturbances. Therefore, to enhance robustness to rapidly changing states, a multi-scale causal convolution module is designed based on time series modeling. The results of time series modeling are processed to explore local patterns and adjust the drone's fine-tuning within a short time window. ReLU activation and batch normalization are used to enhance the response to emergency acceleration and deceleration features, resulting in the following initial neural network: ; Where, is the convolution kernel width, is the number of input channels, is the convolution kernel is the time offset, is the output channel, is the input channel, Bias term; After the structure of the initial neural network is determined, a multi-objective loss function of the initial neural network is constructed to improve the prediction accuracy and dynamic smoothness, and the neural network is obtained; Here, the total loss is composed of the prediction error and the physical constraint: ; Where, is the total loss, is the weight coefficient of the trajectory prediction loss term, is the predicted future time step, is the index of the current time step in the future sequence, For the The true target value of time steps, For the The predicted value of the time step, is the weight coefficient of the smoothing loss term, is the smoothness loss, which is used to constrain the continuity and smoothness of the predicted trajectory; The dataset is used as the input of the neural network, that is, based on a large amount of historical data collected previously, the neural network is trained offline to achieve efficient online decision-making in multi-UAV collaborative tasks; specifically, an independent policy network (Actor) is constructed for each UAV, which is used to map high-dimensional observation input to control action output. The training process adopts a supervised learning paradigm, and by minimizing the difference between the policy network's predicted actions and the actual expert actions, the network parameters are continuously updated to improve the generalization performance and robustness of the strategy; in order to facilitate network training, the training data of each UAV is batch loaded. Specifically, for each UAV, a custom dataset is constructed using the pre-processed observation data and action data to ensure that the data dimension meets the training requirements. Subsequently, the DataLoader function provided by PyTorch is used to randomly shuffle and batch divide the dataset. The batch size is determined by the hyperparameter batch. size is specified, which not only improves data utilization efficiency but also helps reduce the noise caused by data correlation during the gradient descent process. An independent Actor network is constructed for each drone. The input of the network is the observation vector (position, attitude, speed), and the output is the control command (thrust, angular velocity). The number of hidden layer units, activation function and learning rate in the network structure are all hyperparameters. The training goal is to make the action output by the network as close as possible to the optimal action collected in the historical record, so that the policy network can accurately predict the optimal control input of the drone in the current state. During the network training process, the mean squared error (MSE) is used as the loss function, and the network parameters are updated through the backpropagation algorithm and gradient descent method. Within each training batch, the train batch method of the Actor network is used to calculate the loss of the current batch and the loss is accumulated and recorded. During the entire training process, the system will periodically output the average loss value of each epoch and save the network parameters under the current training round for subsequent policy initialization and online control. Get the trained neural network: ; Where, To predict the future The output of the step, is the feature fusion function, is the feature extraction function of Transformer, is the feature extraction function of the long short-term memory network, is the feature extraction function of CNN, From time 1 to The input sequence of The trained neural network is embedded into the model predictive control framework to add the status and quantity of other drones to the model predictive control framework of the current drone, thereby obtaining a distributed model predictive control framework for the current drone. Based on the distributed model predictive control frameworks of all drones, a distributed model predictive control framework for multiple drones is formed. According to the distributed model predictive control framework of the current UAV, the model of the crossing box and the dynamic model of the current UAV, the path prediction planning is performed for the current UAV and other UAVs, and the action of the current UAV is output; the action is executed according to the current state of the current UAV, that is, the state of the current UAV at the next moment is obtained; all UAVs are traversed to obtain the state of all UAVs at the next moment; at the next moment, the action and state of the current UAV are still output according to the state of other UAVs, and the state of all UAVs is traversed; all moments are traversed to obtain the real-time position of each UAV.

[0045] In this step, a dual-modal trajectory predictor is designed to further improve the accuracy and safety of cooperative control. Specifically, when historical state data is insufficient, a constant velocity model is used for simplified kinematic prediction. When data is sufficient, a pre-trained Transformer-LSTM-CNN hybrid model is used to predict the future state of neighboring drones. This prediction module not only improves the accuracy of neighboring drone state estimation, but also provides more accurate cooperative constraint parameters for local optimization problems, thereby effectively reducing potential collision risks. The architecture of the proposed hybrid model, such as Figure 2 As shown in the figure, it includes the following connected in sequence: input end, Transformer encoder, LSTM module, CNN module and output end. The Transformer encoder includes the following connected in sequence: Multi-Head Attention module, Add&Norm module, FeedForward module and Add&Norm module. The architecture consists of three core modules: 1) Transformer encoder captures global spatiotemporal dependencies; 2) LSTM network models temporal continuity; 3) CNN extracts local motion patterns; finally, multimodal feature integration is achieved through the attention weighted fusion layer to output multi-step prediction results; the design is optimized for the nonlinear and strongly coupled characteristics of the drone trajectory, and the expression is: ; Where, To predict the future The output of the step, is the feature fusion function, is the feature extraction function of Transformer, is the feature extraction function of LSTM, is the feature extraction function of CNN, From time 1 to The input sequence is the historical state information.

[0046] Step 106: Perform multi-UAV collaborative crossing based on the real-time position of each UAV.

[0047] In this step, how to perform multi-UAV collaborative crossing based on the real-time position of each UAV belongs to the existing technology and will not be described in detail here.

[0048] In this embodiment, if Figure 3 In the crossing scene shown, the circle represents the starting point and the triangle represents the end point. represents the length of the pendulum, that is, the distance from the suspension point to the center of the through-frame, is the inner ring width of the crossing frame, is the inner ring height of the crossing frame.

[0049] Learning-based distributed model predictive control combines distributed MPC with data-driven policy training to achieve real-time solution and adaptive adjustment of complex dynamic constraints in multi-UAV collaborative missions. This framework not only meets real-time and safety requirements, but also continuously optimizes online decision-making strategies through offline data learning, providing a practical solution for multi-UAV system control in complex environments.

[0050] This application proposes a distributed model predictive control (DMPC) method for multi-UAV collaboration. To effectively avoid collisions between UAVs while ensuring control performance, a deep learning-based trajectory prediction network for neighboring UAVs is introduced into the system. The overall framework can be summarized as follows: when executing distributed MPC, each UAV needs to obtain or estimate the motion state of neighboring UAVs in real time for several steps in the future and use this information as dynamic obstacles or coupling constraints in its local optimization problem. At the same time, a hybrid Transformer-LSTM-CNN prediction model is used to capture the spatiotemporal dependencies between multiple UAVs, outputting more accurate neighbor trajectory prediction results, thereby reducing collisions and coupling conflicts.

[0051] The main processes of the framework include: 1) Multi-UAV state acquisition and local MPC optimization: Each UAV obtains state observations of itself and neighboring UAVs, and generates a control sequence for the next moment or several moments based on the rolling optimization concept of DMPC. DMPC updates its respective control decisions through iterative information exchange, ensuring good global coordination and convergence in distributed scenarios. 2) Prediction of the future trajectory of neighboring UAVs: In order to accurately estimate the possible position and speed changes of neighboring UAVs in the next period of time, a deep network is used to make short-term predictions of the motion state of neighboring UAVs. Unlike the traditional simple constant speed / constant acceleration assumption, this application uses Transformer to handle global temporal dependencies, LSTM to focus on short-term dynamic features, and CNN for local convolution extraction or multi-feature fusion, which improves prediction accuracy and robustness. 3) Coupling Constraints and Collision Avoidance Mechanism: The predicted neighboring aircraft trajectories are treated as coupling terms in the distributed MPC. Safety distance constraints and collision penalty terms are added to the optimization objective or cost function to ensure that the UAVs will not collide in the future time domain. In addition, to address model uncertainty or prediction errors, a safety margin (safety distance scaling) is introduced in the DMPC to further ensure system safety.

[0052] The above-mentioned multi-UAV collaborative crossing method is a trajectory optimization method based on model predictive control (MPC). Through spatiotemporal decoupling modeling (multi-head attention), dynamic gated fusion (Transformer-LSTM-CNN) and physical constraint injection, it achieves high precision and strong robustness in UAV trajectory prediction, while meeting the dynamic constraint requirements of the actual flight control system. It provides a reliable state estimation basis for the autonomous navigation system, enabling multiple UAVs to complete the crossing task safely and efficiently under the constraints of dynamic and static obstacles. It is especially suitable for the collaborative crossing scenario of multiple UAVs under the constraints of dynamic and static obstacles.

[0053] To achieve collaborative modeling of global features and local memory, a trajectory prediction and optimization method for multi-UAV dynamic traversal scenarios was designed. The overall framework is based on model predictive control (MPC). A multi-head self-attention mechanism is introduced to achieve spatiotemporal decoupled modeling to capture the complex interactions between multiple UAV trajectories. Furthermore, the global representation capabilities of the Transformer are combined with the temporal memory of the LSTM, and multi-scale convolutional modules are incorporated to enhance the perception of short-term trajectory perturbations. This results in a deep fusion architecture with dynamic gating capabilities. This architecture not only accurately predicts the future states of other UAVs but also ensures that the predictions meet the execution requirements of the actual flight control system by introducing physical feasibility constraints (such as dynamic limits and control boundaries). Furthermore, the loss function design comprehensively considers trajectory prediction error and smoothness requirements, achieving a multi-objective optimization approach to improve both trajectory accuracy and system stability. Thanks to this design, the proposed method enables collaborative state estimation and high-quality trajectory generation among multiple UAVs in complex environments with both dynamic and static obstacles, ensuring a good balance between safety and efficiency, significantly improving the overall performance of the multi-UAV system in traversal missions.

[0054] It should be understood that although Figure 1 The steps in the flowchart are shown in sequence as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified in this document, there is no strict order restriction for the execution of these steps, and these steps can be executed in other orders. In addition, Figure 1 At least part of the steps may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily executed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least part of the sub-steps or stages of other steps.

[0055] This application also provides a multi-UAV cooperative crossing device, such as Figure 4 As shown, in one embodiment, it includes: a first module 401, a second module 402, a third module 403, a fourth module 404, a fifth module 405 and a sixth module 406, wherein: The first module 401 is used to obtain scene information of multi-UAV collaborative traversal, collect information, and generate a data set; The second module 402 is used to model the traversal frame based on the scene information of the multi-UAV collaborative traversal to obtain a traversal frame model; The third module 403 is used to model each UAV based on the scene information of the multi-UAV cooperative crossing to obtain a dynamic model of each UAV; The fourth module 404 is used to build a model predictive control framework for each UAV based on the dynamic model of each UAV and in combination with a quadratic programming optimization method; The fifth module 405 is configured to obtain the historical trajectory sequence of each UAV, construct a neural network, and train the neural network using the dataset as input to obtain a trained neural network; embed the trained neural network into a model predictive control framework to form a distributed model predictive control framework for multiple UAVs; perform path prediction planning for the current UAV and other UAVs based on the distributed model predictive control framework of the current UAV, the model of the traversal frame, and the dynamic model of the current UAV, output the current UAV's motion, and traverse all UAVs to obtain the real-time position of each UAV; The sixth module 406 is used to perform multi-UAV collaborative crossing based on the real-time position of each UAV.

[0056] For the specific definition of a multi-UAV cooperative traversal device, please refer to the definition of a multi-UAV cooperative traversal method above, and will not be repeated here. Each module in the above-mentioned device can be implemented in whole or in part by software, hardware, or a combination thereof. Each of the above-mentioned modules can be embedded in or independent of the processor in the computer device in hardware form, or can be stored in the memory of the computer device in software form, so that the processor can call and execute the operations corresponding to each of the above modules.

[0057] The contents not described in detail in this specification belong to the prior art known to those skilled in the art.

[0058] The technical features of the above embodiments can be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0059] The above-described embodiments merely represent several implementation methods of the present application. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present application. It should be noted that a person of ordinary skill in the art may make various modifications and improvements without departing from the spirit of the present application, and these modifications and improvements fall within the scope of protection of the present application. Therefore, the scope of protection of the present application shall be based on the attached application documents.

Claims

1. A multi-UAV collaborative crossing method, characterized in that: include: Obtain scene information of multi-UAV collaborative traversal, collect information, and generate data sets; According to the scene information of multi-UAV collaborative crossing, the crossing frame is modeled to obtain the crossing frame model; Based on the scene information of multi-UAV collaborative crossing, each UAV is modeled to obtain the dynamic model of each UAV; Based on the dynamic model of each UAV and combined with the quadratic programming optimization method, a model predictive control framework for each UAV is constructed; The historical trajectory sequence of each drone is obtained, and a neural network is constructed. The neural network is trained using the dataset as input to obtain a trained neural network. The trained neural network is embedded in the model predictive control framework to form a distributed model predictive control framework for multiple drones. Based on the distributed model predictive control framework of the current drone, the model of the traversal box, and the dynamic model of the current drone, path prediction and planning are performed for the current drone and other drones, the current drone's actions are output, and all drones are traversed to obtain the real-time position of each drone. Based on the real-time position of each drone, multi-drone collaborative crossing is carried out.

2. The multi-UAV cooperative crossing method according to claim 1, characterized in that: Obtain scene information of multi-UAV collaborative traversal, collect information, and generate a data set, including: Obtain scene information of multi-UAV collaborative traversal and obtain the initial environment state; Initial observation information and initial state information are obtained based on the initial environment state; a corresponding Gaussian strategy is constructed for each drone based on the initial observation information and initial state information; random sampling is performed in each drone's Gaussian strategy to obtain each drone's initial action and form a set of candidate actions; the candidate actions interact with the environment to obtain reward indicators and pass indicators; based on the reward indicators and pass indicators, the sample weights are calculated and the Gaussian strategy of each drone is updated; new candidate actions are formed based on the updated strategy, and the new candidate actions, initial environment state, initial observation information, and initial state information are used as data at the current sampling time; Update the environment state and send the new candidate action to the environment to drive the system update, obtain the next environment state, the next observation information, and the next state information; obtain the next candidate action based on the next observation information and the next state information, and use the next candidate action, the next environment state, the next observation information, and the next state information as the data for the next sampling time; The environmental status is updated cyclically until the end of the sampling period, and a data set is generated based on the data at all sampling moments.

3. A multi-UAV cooperative crossing method according to claim 1 or 2, characterized in that: According to the scene information of multi-UAV collaborative crossing, the crossing frame is modeled to obtain the crossing frame model, including: Based on the scene information of multi-UAV collaborative crossing, the state vector of the crossing box is defined; Perform force analysis on the crossing frame and obtain the change in the swing angle of the crossing frame based on the state vector of the crossing frame; According to the change of the swing angle of the crossing frame, the expression of the center position of the crossing frame is obtained; Based on the expression of the center position of the crossing frame, the expression of the corner point position of the crossing frame is obtained; According to the expression of the corner point position of the crossing frame, the center position of the crossing frame is obtained and used as the model of the crossing frame.

4. A multi-UAV cooperative crossing method according to claim 1 or 2, characterized in that: Based on the scene information of multi-UAV collaborative crossing, each UAV is modeled to obtain the dynamic model of each UAV, including: According to the scene information of multi-UAV collaborative traversal, the state vector of each UAV is obtained; According to the state vector of each UAV, under the constraint of the objective function, the control vector of each UAV is obtained; According to the control vector of each UAV, the linear velocity change rate and attitude change rate of each UAV are obtained; According to the linear velocity change rate and attitude change rate of each UAV, dynamic modeling is performed on each UAV to obtain the dynamic model of each UAV.

5. The multi-UAV cooperative crossing method according to claim 4, characterized in that: The objective function includes: in, Where, is the objective function, is the target approach term, is the trajectory tracking term, To control the smoothing term, is the collaborative collision avoidance term, For the The state of the moment, is the desired end state, is a positive definite matrix, For the time, To predict the total number of steps, is the time weighting factor, For the The first drone The predicted state of the step, For the The reference state of the step, is the state error weight matrix, For the The drone The control input of the step, is the control input in the hover state, is the control weight matrix, is the predicted number of steps of other drones, For the drones, For the drones, is the collision avoidance gain parameter, is the current position of the drone, For the The drone The position of the step, Is a positive number.

6. A multi-UAV cooperative crossing method according to claim 1 or 2, characterized in that: Based on the dynamic model of each UAV and combined with the quadratic programming optimization method, a model predictive control framework for each UAV is constructed, including: Where, is the dynamic model of the UAV, is the state vector of the UAV, is the position coordinate of the UAV in the inertial coordinate system, is the attitude quaternion corresponding to the drone, is the linear velocity component of the UAV in the inertial coordinate system, is the control vector of the UAV; is the total thrust in the body coordinate system, is the angular velocity component in the body coordinate system; Perform Euler discretization: Where, For drones The state of the moment, For drones The state of the moment, To control the time step, For control input.

7. A multi-UAV cooperative crossing method according to claim 1 or 2, characterized in that: Obtain the historical trajectory sequence of each drone and build a neural network, including: Obtain the historical trajectory sequence and map it to a high-dimensional feature space through a linear embedding layer to obtain trajectory data; Design an encoder and stack it with a multi-head self-attention mechanism and a feedforward network to model the spatiotemporal dependencies in the trajectory data and obtain the encoder output; The output of the encoder is used as the input of the residual enhanced long short-term memory network for temporal modeling; Design a multi-scale causal convolution module to process the results of time series modeling to adjust the drone's fine-tuning within a short time window and obtain the initial neural network; Construct a multi-objective loss function of the initial neural network and obtain the neural network.

8. A multi-UAV cooperative crossing method according to claim 1 or 2, characterized in that: The dataset is used as the input of the neural network to train the neural network and obtain the trained neural network, including: Where, To predict the future The output of the step, is the feature fusion function, is the feature extraction function of Transformer, is the feature extraction function of the long short-term memory network, is the feature extraction function of CNN, From time 1 to The input sequence of .

9. A multi-UAV cooperative crossing method according to claim 1 or 2, characterized in that: The trained neural network is embedded into the model predictive control framework to form a distributed model predictive control framework for multiple drones, including: The trained neural network is embedded into the model predictive control framework to add the status and quantity of other drones to the model predictive control framework of the current drone, thus obtaining a distributed model predictive control framework for the current drone. Based on the distributed model predictive control framework of all drones, a distributed model predictive control framework for multiple drones is formed.

10. A multi-UAV cooperative crossing device, characterized in that: include: The first module is used to obtain scene information of multi-UAV collaborative traversal, collect information, and generate a data set; The second module is used to model the crossing frame based on the scene information of the multi-UAV collaborative crossing, and obtain the crossing frame model; The third module is used to model each UAV based on the scene information of multi-UAV collaborative crossing and obtain the dynamic model of each UAV; The fourth module is used to build a model predictive control framework for each UAV based on its dynamic model and combined with quadratic programming optimization method; The fifth module is used to obtain the historical trajectory sequence of each drone, construct a neural network, and use the dataset as the input to train the neural network to obtain a trained neural network. The trained neural network is embedded in the model predictive control framework to form a distributed model predictive control framework for multiple drones. Based on the distributed model predictive control framework of the current drone, the model of the traversal box, and the dynamic model of the current drone, the path prediction planning is performed for the current drone and other drones, the action of the current drone is output, and all drones are traversed to obtain the real-time position of each drone. The sixth module is used to conduct multi-UAV collaborative crossing based on the real-time position of each UAV.

Images