A human-robot-heterogeneous unmanned aerial vehicle cooperative active perception uncertainty minimization method

Abstract

The application discloses a manned-unmanned aerial vehicle cooperative active sensing uncertainty minimization method, comprising the following steps: constructing a cooperative active sensing system containing manned vehicles and unmanned aerial vehicles; establishing a continuous motor level dynamics model for each unmanned aerial vehicle, discretizing and constructing a discrete prediction model, measurement noise covariance and continuous field of view weight; performing cooperative Kalman filtering based on the measurement noise covariance and the continuous field of view weight, and propagating the target state estimation and the estimation covariance online; reconstructing a prediction control target function containing an estimation covariance trace accumulation item and a constraint set; solving and outputting the control quantity online by using a sampling type prediction controller; and forming a closed loop of state sensing, cooperative estimation and prediction control in each control period of each unmanned aerial vehicle. The application embeds the estimation covariance trace as an uncertainty measure into a prediction control cost function, drives the unmanned aerial vehicle to actively select a flight trajectory for reducing the estimation uncertainty, and is suitable for various civil scenarios.

Description

Technical Field

[0001] This application belongs to the field of cooperative active perception and predictive control technology for aircraft, specifically involving a method for minimizing the uncertainty of cooperative active perception between manned aircraft and heterogeneous unmanned aerial vehicles. Background Technology

[0002] With the rapid development of unmanned aerial vehicle (UAV) technology, collaborative systems centered on manned aircraft and assisted by heterogeneous UAVs are showing broad application prospects in tasks such as search and rescue, target localization, and continuous tracking. These systems typically consist of a manned aircraft and multiple multi-rotor UAVs of varying architectures. The manned aircraft is responsible for macro-level situational awareness and mission guidance, while the UAVs perform detailed detection and target tracking tasks. In the process of collaborative active perception, optimizing UAV flight trajectories to reduce the uncertainty in target state estimation, while ensuring the safe and feasible flight of each UAV in complex environments, is key to improving the overall system effectiveness.

[0003] In existing technologies, manned-unmanned collaborative systems typically employ a hierarchical or centralized control architecture, where the ground station or manned aircraft centrally plans the UAV trajectory, with the UAV merely acting as a command execution terminal. Regarding active perception, existing methods reduce estimation uncertainty by constructing measurement quality indices related to the target's geometric relationship and incorporating them into the trajectory optimization cost function. For UAV control, rigid body dynamics-based control methods are commonly used, employing desired forces and torques as control commands, which are then calculated and distributed to each rotor via a low-level distributor. In constraint handling, actuator upper and lower bounds are typically used as input limits, and these are pruned externally to the controller. Regarding the integration of estimation and control, existing research uses Kalman filters to estimate the target state and incorporates covariance information as feedback for trajectory optimization, forming a closed loop between perception and control.

[0004] However, existing technologies still have the following shortcomings: First, existing collaborative system architectures mostly rely on centralized planning or direct commands from manned aircraft, failing to fully utilize the advantages of autonomous estimation and control by each UAV under a decentralized framework. The systems lack robustness to changes in communication topology and struggle to flexibly adapt to individual differences across heterogeneous platforms. Second, in UAV dynamics modeling, rigid-body models are commonly used, treating desired forces and torques as control commands. This ignores the dynamic process of the thrust state itself and the limitations of the thrust rate of change, leading to the expected commands output by the planning layer potentially failing to be accurately implemented during actual execution due to limited motor response capabilities, thus reducing trajectory executability. Third, in terms of uncertainty optimization for active perception, existing methods typically model measurement effectiveness as discrete gating based on the field-of-view boundary. This gating produces discontinuous jumps when the target enters or exits the field-of-view boundary, resulting in a non-differentiable cost function, which is detrimental to continuous evaluation in gradient-based optimization methods or sampled predictive control. Simultaneously, the modeling of measurement noise covariance fails to fully couple the anisotropic effects of observation distance and line-of-sight direction on measurement accuracy, limiting the physical realism of active perception optimization. Fourth, existing methods lack online verification and anomaly isolation mechanisms for the reliability of manned machine measurements when manned machines participate in collaborative perception. If a manned machine sensor suddenly malfunctions or measurement noise is excessive, the measurement will directly contaminate the entire collaborative estimation result, affecting system stability. Fifth, existing technologies often employ an open-loop approach of "estimating first, then planning" to integrate estimation and control. This fails to jointly predict the evolution of the target state estimation covariance with the UAV's motion within the predictive control framework, making it difficult to achieve truly proactive perception closed-loop optimization. Summary of the Invention

[0005] To address the aforementioned problems in the existing technology, this application provides a method for minimizing uncertainties in manned-heterogeneous unmanned aerial vehicle (UAV) cooperative active perception. The technical problem to be solved by this application is achieved through the following technical solution: A method for minimizing uncertainties in manned-heterogeneous unmanned aerial vehicle (UAV) cooperative active perception includes: S100, Construct a collaborative active perception system, which consists of a manned aircraft and at least one unmanned aircraft, wherein the manned aircraft periodically broadcasts its own status to the unmanned aircraft; S200: For each of the aforementioned UAVs, a continuous motor-level dynamic model is established and discretized based on the sampling period to construct a discrete prediction model; S300, for each of the aforementioned UAVs, based on the UAV pose and target feature state predicted by the discrete prediction model, construct the measurement noise covariance and continuous field of view weights. S400, each of the UAVs performs cooperative Kalman filtering based on the measurement noise covariance and the continuous field of view weight, splicing and weighting the multi-platform measurements including the manned UAV measurements, and propagating the target state estimate and estimated covariance online. S500, each of the UAVs estimates the extended state by combining the estimated covariance with the target state estimate, and establishes a unified discrete propagation form of the extended state in the prediction time domain; S600, each of the UAVs constructs a predictive control objective function and a set of constraints containing the estimated covariance trace based on the extended state and the estimated covariance trace; S700, each of the UAVs uses a sampled predictive controller to solve the problem online based on the discrete prediction model, the objective function, and the constraint set, and outputs the first control quantity of the current control cycle; S800, each of the UAVs uses the extended state as the initial value, predicts its own state using the discrete prediction model, updates the target estimate using the cooperative Kalman filter, and solves the control quantity using the sampled predictive controller. In each control cycle, a closed loop of state perception, cooperative estimation and predictive control is formed, and the extended state and nominal control sequence are updated on a rolling basis.

[0006] Optionally, S100 includes: S110, defines a manned aircraft periodically broadcasting its own navigation status, the navigation status including at least position, speed and attitude information; S120 defines that each UAV operates independently in a distributed communication framework for estimation and control, and the UAVs share measurement information and predicted trajectories through communication, and each UAV receives navigation status broadcast by the manned aircraft; S130, construct a collaborative active perception system for the manned aircraft and various unmanned aircraft.

[0007] Optionally, the S200 includes: S210, for the i-th UAV, define the motor-level state and control input, and set hard constraints on the thrust vector and thrust rate of change; wherein, the motor-level state includes the airframe state and thrust vector, and the control input is the thrust rate of change; the airframe state includes position, attitude quaternion, linear velocity and angular velocity; S220, For the i-th UAV, establish a continuous motor-level dynamics model that includes rigid body dynamics and thrust state dynamics, wherein rigid body dynamics describes the changes in the state of the UAV, and thrust state dynamics describes the changes in the thrust vector with the control input; S230, for the i-th UAV, based on the rotor action point position, thrust direction unit vector, rotation direction sign and anti-torque proportional coefficient, construct the resultant force distribution matrix and resultant torque distribution matrix column by column to uniformly express the continuous motor-level dynamics model of UAVs with different configurations; S240, the continuous motor-level dynamic model is discretized based on the sampling period. In each prediction step, hard constraint limiting is applied to the control input. The dynamic feasible input boundary is calculated based on the thrust hard constraint and the current thrust state to ensure that the thrust state after one integration satisfies the thrust hard constraint. At the same time, the attitude quaternion is normalized to construct a discrete prediction model.

[0008] Optional, the S300 includes: S310, For the i-th UAV, based on the UAV pose predicted by the discrete prediction model and the sensor fixed external parameters, calculate the sensor pose in the world coordinate system, extract the sensor principal axis direction, and combine the position in the target feature state to calculate the relative distance and line of sight between the sensor and the target. S320, construct a measurement noise covariance ellipsoid model coupled with the relative distance and the line of sight direction, so that the measurement uncertainty is anisotropic along the line of sight axis and the transverse tangent plane, and introduce an optimal observation distance mechanism so that the measurement noise covariance increases as the observation distance deviates from the optimal observation distance; S330, based on the cosine of the angle between the sensor's half-angle of the field of view and the deviation of the line of sight from the sensor's main axis, the continuous field of view weight is constructed using the Sigmoid function, so that the continuous field of view weight smoothly approaches one as the target enters the field of view and smoothly approaches zero as the target leaves the field of view.

[0009] Optional, the S400 includes: S410, each of the UAVs splices together the measurements, observation matrices, measurement noise covariance and continuous field of view weights of each observation source in a block diagonal form, and splices together the measurements of the manned aircraft as optional observation sources to form a multi-platform spliced ​​measurement. S420, each of the UAVs performs a reliability determination on the measurements of the manned aircraft, calculates a normalized squared innovation index based on the current target state estimate, and determines an anomaly and performs rapid isolation when the normalized squared innovation index exceeds a preset threshold; otherwise, the measurement covariance of the manned aircraft is equivalently amplified according to a preset reliability weight to obtain an equivalent measurement covariance. S430, each of the UAVs uses a time-continuous Kalman filter equation to propagate the target state mean and estimated covariance, wherein the Kalman gain is calculated from the estimated covariance, the observation matrix and the equivalent measurement covariance, and is updated online in steps of the control period through discretization.

[0010] Optional, the S500 includes: S510, stack the lower triangular elements of the estimated covariance matrix in a fixed order to form a minimum vector representation, and define a reconstruction mapping that reconstructs the estimated covariance matrix from the minimum vector representation; S520, each of the UAVs will combine the motor-level state, the target state estimate, and the minimum vector representation to form an extended state, wherein the motor-level state includes at least the body state and the thrust vector; S530, each of the UAVs establishes a unified discrete progression form of the extended state in the prediction time domain based on the discrete prediction model and the discrete recursive form of the cooperative Kalman filter, so that the extended state is updated synchronously in a fixed recursive order at each prediction step.

[0011] Optional, the S600 includes: S610, each of the UAVs constructs a total cost function in the prediction time domain, the total cost function including the weighted sum of squared output errors obtained from the extended state mapping, the trace weighted cumulative term calculated from the estimated covariance, and the control input regularization term; S620, each of the aforementioned UAVs is subject to a set of constraints, which includes extended dynamic constraints based on a unified discrete propulsion form, hard constraints on thrust and thrust change rate, and unified safety interval constraints between UAVs and between UAVs and manned aircraft. The unified safety interval constraints adopt an axis-aligned safety box form and are determined by an intrusion function. When the intrusion function is greater than zero, it indicates that the UAV has entered the safety box region.

[0012] Optional, the S700 includes: S710, each of the UAVs maintains a nominal control sequence, which is the control input sequence in the prediction time domain obtained by left shifting after the previous control cycle update, and the nominal control sequence is used as the reference sequence of the current cycle sampling predictive controller; S720, each of the UAVs generates a candidate control sequence by adding sampling noise based on the nominal control sequence, and sequentially performs hard constraint limiting and dynamic feasible input boundary pre-verification induced by the thrust hard constraint and the current thrust state on the candidate control input; S730, each of the UAVs takes the current extended state as the initial value and performs forward prediction on each sampled trajectory based on the discrete prediction model. During the prediction process, the total cost is accumulated based on the objective function and the constraint set. S740, each of the UAVs calculates the exponential weight of all sampled trajectories. The exponential weight is calculated using an exponential function based on the difference between the total cost of each sampled trajectory and the minimum total cost. The nominal control sequence is then weighted and corrected according to the exponential weight. The first control quantity of the corrected sequence is output, and the corrected nominal control sequence is used as the initial maintenance value for the next control cycle.

[0013] Optionally, the S800 includes: S810, in each control cycle, each UAV uses the current extended state as the initial value, predicts its own state sequence in the future time domain based on the discrete prediction model, and updates the target state estimate and estimated covariance based on the multi-platform measurements obtained in the current cycle using the cooperative Kalman filter. S820: The updated target state estimate and estimated covariance are embedded into the extended state. The control quantity for the current cycle is solved and executed based on the sampled predictive controller. The updated extended state and the left-shifted nominal control sequence are used as the initial values ​​for the next control cycle, forming a closed loop of state perception, cooperative estimation and predictive control.

[0014] Optionally, the estimated covariance trace is used as an uncertainty measure, and its cumulative term is embedded in the predictive control objective function. Through the evaluation and weighted update of the candidate trajectory by the sampled predictive controller, each UAV is driven to select a flight trajectory that can reduce the estimated covariance trace under the premise of satisfying the continuous motor-level dynamics model and the constraint set, so as to achieve the goal of minimizing the uncertainty of collaborative active perception.

[0015] Beneficial effects: 1. This application constructs a collaborative active perception system composed of manned aircraft and heterogeneous multi-rotor UAVs. It adopts a distributed communication framework to enable each UAV to operate independently for estimation and control, overcoming the shortcomings of the existing centralized architecture in terms of insufficient robustness to changes in communication topology. At the same time, by constructing the resultant force distribution matrix and resultant torque distribution matrix column by column based on the rotor action point position, thrust direction unit vector, rotation sign, and anti-torque proportional coefficient, it achieves a unified expression of the dynamic model of UAVs with different configurations such as quadcopters, hexacopters, and tiltrotors. It can flexibly adapt to the individual differences of heterogeneous platforms without changing the dynamic expression form, significantly improving the scalability and adaptability of the system.

[0016] 2. This application establishes a motor-level dynamics model that includes rigid body dynamics and thrust state dynamics, using the thrust change rate as the control input. In each prediction step, it calculates the dynamic feasible input boundary based on the thrust hard constraint and the current thrust state, ensuring that the thrust state after one integration satisfies the thrust hard constraint. This solves the problem that existing rigid body-level models ignore the thrust dynamic process and only handle actuator constraints through external pruning, resulting in the difficulty in accurately executing planning instructions. At the same time, by constructing a measurement noise covariance ellipsoid model coupled with the observation distance and line of sight, the measurement uncertainty exhibits anisotropy along the line of sight axis and the lateral tangent plane. Furthermore, by using the Sigmoid function to construct continuous field of view weights, it overcomes the defect of discontinuous jumps in the cost function caused by discrete gating, making the field of view gating continuously change in the prediction, thus improving the physical realism and numerical stability of active perception optimization.

[0017] 3. This application introduces a reliability judgment and anomaly isolation mechanism for manned machine measurements into the cooperative Kalman filter. Based on the normalized innovation square index, it performs consistency checks on manned machine measurements. When the index exceeds the threshold, it performs rapid isolation and amplifies the measurement covariance according to the reliability weight. This effectively suppresses the contamination of cooperative estimation by abnormal or unreliable measurements, solving the problem of lack of protection against sudden anomalies of manned machine sensors in existing methods. By embedding the minimum vector representation of the estimated covariance and the target state estimation into the extended state and advancing it in a unified discrete form in the prediction time domain, it realizes the joint prediction of the estimated covariance evolution and UAV motion. This overcomes the limitations of the existing "first estimate, then plan" open-loop method. Through the evaluation and weighted update of candidate trajectories by the sampling predictive controller, it drives the UAV to select a flight trajectory that can reduce the estimated covariance trajectory under the premise of satisfying dynamics and constraints, truly realizing closed-loop optimization of active perception.

[0018] The present application will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0019] Figure 1 This is a flowchart illustrating a method for minimizing uncertainties in a manned-heterogeneous unmanned aerial vehicle (UAV) cooperative active perception method provided in this application. Figure 2 This is a schematic diagram of the process loop of the manned-heterogeneous unmanned aerial vehicle cooperative active perception uncertainty minimization method provided in this application. Detailed Implementation

[0020] The present application will be described in further detail below with reference to specific embodiments, but the implementation of the present application is not limited thereto.

[0021] This application belongs to the field of cooperative active perception and predictive control technology for aircraft, and proposes a method for minimizing the uncertainty of cooperative active perception between manned aircraft and heterogeneous unmanned aerial vehicles (UAVs) based on motor-level dynamics and sampled predictive control. The system consists of a manned aircraft (MAV) and multiple heterogeneous multi-rotor UAVs. The MAV flies freely and periodically broadcasts its own state (optionally broadcasting target observations). The UAVs independently complete estimation and control within a distributed communication framework. To address the UAV executability problem, motor-level nonlinear dynamics modeling of the airframe state and thrust state is adopted, with the thrust change rate as the control input. The explicit satisfaction of actuator constraints is achieved through a dynamic feasible input bound induced by thrust constraints. For cooperative active perception, multi-platform measurement stitching and continuous field-of-view weights are constructed. The validity of intermittent observations is embedded into a cooperative Kalman-Bucy filter in the form of continuous weights to achieve online propagation of target state and estimated covariance. At the same time, a MAV measurement reliability verification and isolation mechanism is introduced to suppress the contamination of cooperative estimation by abnormal measurements. Using covariance trace as an uncertainty measure, a rolling time-domain cost is formed by combining motion task errors. Sampled predictive control (MPPI) is then employed to randomly perturb the nominal control sequence, perform forward prediction, and exponentially weighted updates, outputting the first control variable to achieve online closed-loop control. This method enables collaborative active perception with multiple heterogeneous UAVs by utilizing only the periodically broadcast navigation status of a manned aircraft while maintaining autonomous flight control, and optimizes uncertainty by minimizing the estimated covariance trace. This method improves motor-level feasibility and adaptability to nonlinear / non-convex fields of view and safety constraints, making it suitable for civilian scenarios such as search and rescue, target localization, and continuous tracking.

[0022] Combination Figure 1 and Figure 2 This application provides a method for minimizing uncertainties in manned-heterogeneous unmanned aerial vehicle (UAV) cooperative active perception, including: S100, Construct a collaborative active perception system, which consists of a manned aircraft and at least one unmanned aircraft, wherein the manned aircraft periodically broadcasts its own status to the unmanned aircraft; S200: For each of the aforementioned UAVs, a continuous motor-level dynamic model is established and discretized based on the sampling period to construct a discrete prediction model; S300, for each of the aforementioned UAVs, based on the UAV pose and target feature state predicted by the discrete prediction model, construct the measurement noise covariance and continuous field of view weights. S400, each of the UAVs performs cooperative Kalman filtering based on the measurement noise covariance and the continuous field of view weight, splicing and weighting the multi-platform measurements including the manned UAV measurements, and propagating the target state estimate and estimated covariance online. S500, each of the UAVs estimates the extended state by combining the estimated covariance with the target state estimate, and establishes a unified discrete propagation form of the extended state in the prediction time domain; S600, each of the UAVs constructs a predictive control objective function and a constraint set containing the estimated covariance trace based on the extended state and the estimated covariance. The estimated covariance trace serves as an uncertainty measure, and its cumulative term is embedded in the predictive control objective function. Through the evaluation and weighted update of candidate trajectories by the sampled predictive controller, each of the UAVs is driven to select a flight trajectory that can reduce the estimated covariance trace, under the premise of satisfying the continuous motor-level dynamics model and the constraint set, so as to achieve the goal of minimizing the uncertainty of collaborative active perception.

[0023] S700, each of the UAVs uses a sampled predictive controller to solve the problem online based on the discrete prediction model, the objective function, and the constraint set, and outputs the first control quantity of the current control cycle; S800, each of the UAVs uses the extended state as the initial value, predicts its own state using the discrete prediction model, updates the target estimate using the cooperative Kalman filter, and solves the control quantity using the sampled predictive controller. In each control cycle, a closed loop of state perception, cooperative estimation and predictive control is formed, and the extended state and nominal control sequence are updated on a rolling basis.

[0024] In one specific embodiment of this application, S100 includes: S110, defines a manned aircraft periodically broadcasting its own navigation status, the navigation status including at least position, speed and attitude information; S120 defines that each UAV operates independently in a distributed communication framework for estimation and control, and the UAVs share measurement information and predicted trajectories through communication, and each UAV receives navigation status broadcast by the manned aircraft; S130, construct a collaborative active perception system for the manned aircraft and various unmanned aircraft.

[0025] Consider using a manned MAV and A cooperative active perception system consisting of heterogeneous multi-rotor unmanned aerial vehicles (UAVs) is proposed. The motion of the MAVs is independently generated by the pilot or their onboard flight control system, and this method does not impose any restrictions on the MAV control law. The MAVs are used to periodically broadcast their own state information to each UAV, and when the MAVs have the ability to observe target features, their observation information can be used as an optional observation source and incorporated into the cooperative estimation process.

[0026] No. UAV ( Different configurations are allowed, resulting in differences in the number of actuators, force / torque distribution, and sensor parameters. To characterize heterogeneity, the first... The inherent parameter set of the UAV is in: For quality; The inertia matrix; Number of rotors (or equivalent channels); and These are the distribution matrices for the resultant force and the resultant moment, respectively. For the field of view; For measuring frequency or equivalent measurement period.

[0027] Establish a world inertial coordinate system For the first Establishing a body coordinate system using a UAV ,use Indicates the body reference point at The position in the middle, using Indicator System Compared to the World System The unit quaternion attitude. Establish a sensor coordinate system for the airborne sensor. It is stipulated that the sensor is fixedly connected to the machine body, and its fixed pose transformation is constant. Let be the fixed pose transformation matrix between the sensor system and the machine system, which remains unchanged during flight. For the ... The UAV defines the motor-level state, input, and hard constraint interface as follows: in: Linear velocity expressed in the world system; Angular velocity expressed by the machine system; This is the thrust vector; Input the thrust rate of change. Thrust and thrust rate of change satisfy a hard constraint. in: This refers to the upper and lower limits of the thrust. These are the upper and lower bounds of the rate of change of thrust; the inequality should be understood element by element according to its components.

[0028] Define the characteristics of the observed target as Its state is A linear Gaussian motion model is adopted. in: This is the state transition matrix; Let be the process noise covariance matrix.

[0029] For the A linear Gaussian observation model was established using one sensor. in: For measurement vectors; The observation matrix; To measure the noise covariance matrix, a field-of-view gating variable is introduced. This indicates the validity of the measurement, and a continuous approximation will be applied subsequently. Coupled with relative geometry (distance and angle) to support active sensing optimization.

[0030] A distributed collaborative structure is adopted: each UAV independently operates its local control and estimation, and shares necessary measurement information through communication to construct the required measurements for the entire team's splicing. Let the control cycle be... The predicted time domain length is The number of prediction steps is ,satisfy in: To control the update cycle; For rolling prediction in the time domain; To predict the number of steps. Perform a rolling update every cycle, output and execute the first control variable.

[0031] In one specific embodiment of this application, S200 includes: S210, for the i-th UAV, define the motor-level state and control input, and set hard constraints on the thrust vector and thrust rate of change; wherein, the motor-level state includes the airframe state and thrust vector, and the control input is the thrust rate of change; the airframe state includes position, attitude quaternion, linear velocity and angular velocity; For the The UAV is modeled using a motor-level approach combining rigid body dynamics and thrust state dynamics. The body state and thrust state are respectively... And let the overall state The control input is the thrust change rate. in: For the first The thrust vector of each rotor (or equivalent channel) of the UAV; This is the input for thrust rate of change control.

[0032] To maintain a consistent expression across different configurations, for the first One rotor ( Define its geometric / aerodynamic parameters: rotor action point position vector (In-system representation), thrust direction unit vector (In-system representation), rotation sign The ratio of reverse torque to thrust Among them: For standard coaxial multirotors, there are usually: (Vertical axis of the engine system); for tilting propeller platforms, Determined by the rotor installation tilt angle (which can be obtained from a constant rotation matrix in the installation direction). (Obtained by rotation).

[0033] Order No. The scalar thrust of each rotor is Then the resultant force and resultant torque of the whole machine under the machine system can be written as: Among them: the first item The moment of the corresponding thrust; the second term The equivalent contribution of the rotor anti-torque to the airframe.

[0034] Will Stacking yields a matrix form. The allocation matrix is ​​defined as constructed column-by-column. Therefore, different configurations such as quadcopters, hexacopteres, and tiltrotors are only distinguished by... Different values ​​of reflect its heterogeneity, without needing to change the dynamic expression.

[0035] S220, For the i-th UAV, establish a continuous motor-level dynamics model that includes rigid body dynamics and thrust state dynamics, wherein rigid body dynamics describes the changes in the state of the UAV, and thrust state dynamics describes the changes in the thrust vector with the control input; Based on the above, the first The motor-stage continuous dynamics of a UAV is written as in: From the above and The symbols are given; the rest are consistent with S100.

[0036] S230, for the i-th UAV, based on the rotor action point position, thrust direction unit vector, rotation direction sign and anti-torque proportional coefficient, construct the resultant force distribution matrix and resultant torque distribution matrix column by column to uniformly express the continuous motor-level dynamics model of UAVs with different configurations; To ensure that attitude control is achievable under the motor-level model, this application addresses the issue of... The following basic feasibility conditions must be applied to the deployment of UAVs: Wherein: This condition guarantees that the three-dimensional resultant moment can theoretically be obtained from... This generates a basis for feasible input assignment for roll / pitch / yaw control; if the platform configuration leads to If the controllable torque space is limited, it does not fall within the scope of the heterogeneous platform preferred for implementation in this application.

[0037] Regarding the "force-torque decoupling" relationship, this application does not mandate that all configurations satisfy complete decoupling, but adopts the following unified criterion to describe the coupling properties: Let This represents the subspace of thrust combinations that do not produce a resultant torque. If a non-zero vector exists... Then, this thrust combination can generate an increment of "free resultant force" without changing the resultant torque. in: The existence and direction of the resultant force are determined by the configuration and are used to characterize the degree to which the platform can independently adjust the force while maintaining a constant torque. For a standard coaxial multirotor, the direction of the resultant force is usually constrained by the vertical axis of the airframe; for a tiltrotor platform, the direction of the resultant force can vary. It becomes more general through variation, thus exhibiting stronger coupling characteristics. Because this application directly uses [the technology / method] in predictive control... , Therefore, regardless of whether the platform is in a coupled, partially decoupled, or approximately decoupled state, it can maintain the physical consistency of dynamic predictions under a unified model.

[0038] Furthermore, since the control input is ,when When it is a constant, there is Therefore, the "force-torque coupling / decoupling" attribute is determined by the allocation matrix itself. Using the thrust change rate as input will not change the static coupling structure, but only make the thrust change subject to the dynamic capability constraint of the motor, thereby improving the rigor of the executability description.

[0039] S240, the continuous motor-level dynamic model is discretized based on the sampling period. In each prediction step, hard constraint limiting is applied to the control input. The dynamic feasible input boundary is calculated based on the thrust hard constraint and the current thrust state to ensure that the thrust state after one integration satisfies the thrust hard constraint. At the same time, the attitude quaternion is normalized to construct a discrete prediction model.

[0040] Hard constraint boundaries for thrust and thrust rate of change are obtained through identification or calibration: in: For the upper and lower limits of thrust, These are the upper and lower bounds of the rate of change of thrust; the above inequalities are all interpreted element by element according to their components.

[0041] Let the sampling period be The continuous motor stage dynamics are discretized in the prediction time domain. A zero-order preservation assumption is adopted: within the interval... Internal control input It is a constant. The discrete propulsion model is written as: in This represents the discrete state mapping obtained by applying Euler or numerical integration (e.g., explicit Runge-Kutta) to the continuous rigid body dynamics described in S210–S220.

[0042] Thrust state is updated using discrete integrals: To ensure that the discrete predicted trajectory always satisfies the motor-level hard constraints throughout the entire prediction time domain and maintains the consistency of the discrete dynamics of "control input-thrust state", this application performs a step-by-step analysis of the discrete predicted trajectory. Perform the following consistent feasibility processing: (1) Input hard constraint limit: (2) Upper and lower bounds of dynamic feasible input induced by thrust hard constraints: (3) Feasibility of “next thrust pre-verification” (ensuring that the thrust after one integration step still meets the hard constraints): (4) Discrete update of thrust state: Among them: the above , and All are understood element-wise, based on components. Since S100 and S200 have already guaranteed... Therefore, there is no need to further... Additional pruning is performed to avoid disrupting the consistency of discrete dynamics. Simultaneously, to prevent numerical integration from causing the attitude quaternions to deviate from the unit modulus, normalization is performed after each discrete update. Finally, to facilitate unified use by subsequent solution modules, the dynamics interface in the prediction time domain is written as follows: in: For control sequences, It is a state sequence generated by discrete propagation; and it specifies that... Each call synchronously performs the consistency processing of "input hard constraint amplitude limiting - thrust-induced dynamic feasibility (next thrust pre-verification) - quaternion normalization" to ensure that the predicted trajectory is always physically feasible in terms of motor-level constraints and attitude representation.

[0043] In one specific embodiment of this application, S300 includes: S310, For the i-th UAV, based on the UAV pose predicted by the discrete prediction model and the sensor fixed external parameters, calculate the sensor pose in the world coordinate system, extract the sensor principal axis direction, and combine the position in the target feature state to calculate the relative distance and line of sight between the sensor and the target. To map the "flight configuration" to "measured mass," the sensor's pose in the world coordinate system is first calculated from the UAV's state and fixed extrinsic parameters. Let the first... The UAV's body position is The sensor is fixed to the external parameters of the machine body. Then there is in: For sensor system In the world system Pose transformation under the following conditions; For machine system In the world system Pose transformation under the following conditions; This is a fixed external parameter (constant) of the sensor relative to the machine body.

[0044] Depend on Extract the sensor's position in the world system. The unit direction of the sensor's principal axis (optical axis) in the world frame is extracted from its rotating part. (The main axis direction is based on the axis defined in the sensor coordinate system, for example, taking...) of (Axis). Let the world-system position of the target feature be... (From characteristic state) (Position component extraction in the middle), defining relative geometric quantities Further define the cosine of the angle between the visual axis and the line of sight. in: Reflects the geometry of the measured depth; Determine the "direction of sight"; This describes the degree of deviation of the feature from the sensor's main axis. To avoid Too small a value leads to unstable values; in engineering implementation, a smaller value is preferable. ,in It is a very small positive number.

[0045] S320, construct a measurement noise covariance ellipsoid model coupled with the relative distance and the line of sight direction, so that the measurement uncertainty is anisotropic along the line of sight axis and the transverse tangent plane, and introduce an optimal observation distance mechanism so that the measurement noise covariance increases as the observation distance deviates from the optimal observation distance; In order for predictive control to "use configurational changes to reduce estimation uncertainty", it is necessary to establish and The coupling relationship is explained. Range-and-bearing sensors typically have different accuracies in lateral angle and depth information. Therefore, the measurement uncertainty is represented as an ellipsoid with anisotropy along the "line-of-sight axis / lateral tangent plane," and an "optimal observation distance" mechanism is introduced: when the observation distance deviates from a certain reference distance, the overall uncertainty increases. Correspondingly, the measurement covariance takes the following form: in: This is the scaling factor; To preset the optimal observation distance; This represents the uncertainty parameter in the lateral (azimuth / pitch) direction; The uncertainty parameter is the line-of-sight axis (depth / distance) direction; This is a rotation matrix used to align the principal axes of the covariance ellipsoid to the line of sight.

[0046] matrix The construction satisfies the alignment condition that "the third principal axis of the ellipsoid is along the line of sight". In implementation, a stable orthogonal basis can be constructed: a reference vector is selected. (e.g., priority selection) If with If they are nearly collinear, then take the following: ),structure in: To and Orthogonal unit vectors; It consists of three columns of orthogonal unit vectors, therefore .when If the vector is too small (nearly collinear), another preset vector can be used. To avoid degradation.

[0047] The above The model causes measurement uncertainty to deviate with observation distance. Increase and increase, while through By correctly aligning the uncertainties in the "depth / lateral direction" to the current line of sight in space, a continuous, differentiable (except for degradation points) and computable measurement quality mapping is provided for proactive perception planning.

[0048] S330, based on the cosine of the angle between the sensor's half-angle of the field of view and the deviation of the line of sight from the sensor's main axis, the continuous field of view weight is constructed using the Sigmoid function, so that the continuous field of view weight smoothly approaches one as the target enters the field of view and smoothly approaches zero as the target leaves the field of view.

[0049] Due to discrete gating This can cause discontinuous jumps when the target approaches the field of view boundary, which is detrimental to continuous optimization in rolling prediction. Therefore, field of view gating is made continuous. Given the first... Half-angle of each sensor field of view Define continuous field of view weights in: The steepness coefficient is used to adjust the degree to which soft gating approximates hard gating; when Significantly greater than hour A value close to 1 indicates that the target is stably located within the field of view; when... Less than hour A value close to 0 indicates that the target is moving out of the field of view. This construction makes the "entering / leaving the field of view" process a continuous change in the prediction, thus ensuring that the subsequent cost function based on the uncertainty index continues as the state changes. To enable the "measurement contribution" to be directly reflected in subsequent collaborative estimation and uncertainty propagation, this application uses continuous field-of-view weights to construct a measurement gating weight matrix.

[0050] For the Each observation source (UAV sensor or MAV sensor) defines a gating weight. Among them, UAVs The values ​​are given by the Sigmoid weights of the field of view; MAV's It can be obtained by multiplying the field of view weight and the reliability weight. This is constructed during multi-source stitching. in For the first The dimension of each observation source is measured. This will be further addressed in subsequent Kalman–Bucy updates. Linear weighting is applied to the innovation term and the covariance shrinkage term, so that when At this point, the observation source tends to fail for filtering updates, but it will not introduce discontinuous gating.

[0051] In one specific embodiment of this application, S400 includes: S410, each of the UAVs splices together the measurements, observation matrices, measurement noise covariance and continuous field of view weights of each observation source in a block diagonal form, and splices together the measurements of the manned aircraft as optional observation sources to form a multi-platform spliced ​​measurement. At discrete time (Corresponding control cycle) (), stitching together all platform measurements from the observations into in: For the first Measurements from individual observation sources; This corresponds to the linear observation matrix; To measure covariance; For field of view weights; An identity matrix that matches the measurement dimension of the observation source; This indicates stacking by column. This indicates that the blocks are joined diagonally.

[0052] When MAV lacks observation capability or there is no effective measurement in this period, it is considered as (Or omit the corresponding MAV item directly), so that MAV does not participate in the collaborative update of this cycle.

[0053] S420, each of the UAVs performs a reliability determination on the measurements of the manned aircraft, calculates a normalized squared innovation index based on the current target state estimate, and determines an anomaly and performs rapid isolation when the normalized squared innovation index exceeds a preset threshold; otherwise, the measurement covariance of the manned aircraft is equivalently amplified according to a preset reliability weight to obtain an equivalent measurement covariance. To avoid excessive noise in MAV measurements or contamination of the overall estimate by sudden failures, this application introduces a reliability weight for MAV measurements. Anomaly gating is performed based on consistency checks. This is based on current prior estimates. With covariance Define innovation quantity and innovation covariance as And construct a normalized innovation square index Let the anomaly detection threshold be... Define exception gating variables when If the MAV measurement is determined to be abnormal or unreliable, rapid isolation should be performed: when If the MAV measurement is deemed reliable, then its differences are weighted to construct an equivalent covariance: in: It is a very small positive number, used to avoid dividing the value by zero; Follow This decreases while increasing, thus automatically reducing the impact of MAV measurements during Kalman updates. To suppress access / exit jitter, a hold-and-recovery counting mechanism can be employed: maintaining isolation during anomalies. One cycle; after the maintenance period ends, continuous [processing is required]. Secondary satisfaction Access was restored only after that.

[0054] use , Replacing the corresponding MAV term in S410 yields the equivalent concatenation matrix used for collaborative updates in this cycle: S430, each of the UAVs uses a time-continuous Kalman filter equation to propagate the target state mean and estimated covariance, wherein the Kalman gain is calculated from the estimated covariance, the observation matrix and the equivalent measurement covariance, and is updated online in steps of the control period through discretization.

[0055] Estimating the mean and covariance using the propagation characteristics in a continuous-time form: in: To estimate the mean of the features, To estimate the covariance; These are the parameters of the characteristic motion model; Kalman gain; The weight matrix includes both field-of-view gating and reliability gating. This includes the equivalent measurement covariance scaled to include reliability.

[0056] In engineering implementation, the control cycle is used. Discretizing the above equations (using Euler or numerical integration) yields the discrete recurrence relation: in: The reliability gating of MAV is already included, so when MAV is abnormal or unreliable, its update items will automatically exit without affecting the collaborative updates of other observation sources.

[0057] To ensure that all UAV propagations are applicable to both online filtering updates and MPPI prediction rollout, this application adopts a consistent processing method for innovative items: during online filtering updates, The actual measurements from each observation source at the current moment are stitched together, and then processed according to S420. Then substitute it into S430 to complete. Online updates; when predicting rollout, to avoid explicitly sampling measurement noise in the planning, the desired measurement is taken. The expected value of the innovation item is zero, so the mean mainly propagates through the process item; mean, however, still propagates through the process item. The combination of terms shrinks or grows, achieving a predictable closed loop of "candidate trajectory - field of view gating - measurement quality - uncertainty evolution".

[0058] To embed the uncertainty propagation results into the predictive control cost, this application selects the covariance trace as the uncertainty measure: Uncertainty cost terms are accumulated at each node within the prediction time domain. in: For uncertain weights, This is the predicted step count. This index and its cumulative form are used in the subsequent construction of the S600 objective function, enabling the controller to reduce its tendency to choose a specific path while satisfying the dynamics and constraints. The trajectory is tracked to achieve the goal of minimizing uncertainty in collaborative proactive perception.

[0059] In one specific embodiment of this application, S500 includes: S510, stack the lower triangular elements of the estimated covariance matrix in a fixed order to form a minimum vector representation, and define a reconstruction mapping that reconstructs the estimated covariance matrix from the minimum vector representation; To propagate estimation uncertainties and reduce the extended state dimension in predictive control, this application uses feature estimation covariance. Represented in minimal vector form. The minimal representation dimension is defined as... structure for The lower triangular elements are stacked in a fixed order, using a "row-by-row stacking of the lower triangular elements" approach: Among them: when When the characteristic state is a three-dimensional position, there is correspond The six independent elements.

[0060] Define refactoring mapping :Depend on recover The rule is as follows: fill in the boxes in the stacking order described above. , and then Thus, a symmetric matrix is ​​obtained. To suppress the asymmetry caused by discrete errors, it is preferable to perform a symmetry-based process after reconstruction. and when needed Numerical positive definiteness protection is performed to ensure that subsequent calculations of Kalman gain are consistent with... Stability.

[0061] S520, each of the UAVs will combine the motor-level state, the target state estimate, and the minimum vector representation to form an extended state, wherein the motor-level state includes at least the body state and the thrust vector; For the This application extends predictive control state to UAVs. in: Motor-level states defined for S200; Estimate the mean of the target features; This represents the minimum covariance.

[0062] During the sampling period Under these conditions, the extended state progresses in a uniform discrete form within the prediction time domain: in: For motor-level control input (thrust rate of change); External parameters (used in nodes) Construct splicing measurement related quantities and teammate / manned machine information).

[0063] To ensure consistency in the logic of "predictive implementation - cost assessment - constraint determination", Maintain a fixed recursive order within each step and include at least: (1) Propulsion based on S200 discrete motor stage dynamics and to Perform hard constraint feasibility; the feasibility includes thrust hard constraints. Induced dynamic feasible input upper and lower bounds To ensure that the points are obtained Automatically satisfy thrust hard constraints, while performing attitude quaternion normalization; (2) By Reconstruct the covariance matrix and propagate it using the discrete intermittent measurement method of S430, specifically the Kalman–Bucy recursive propagation. and and will Compress it again ; (3) During the propagation process, the covariance numerical consistency processing (symmetry and positive definiteness protection) can be optionally performed to ensure the stability of gain calculation and uncertainty index calculation in the prediction time domain.

[0064] The above "will The organization of "embedding extended states and propagating them in the rolling time domain" enables predictive control to explicitly utilize the evolution of uncertainty to achieve proactive sensing closed loops.

[0065] S530, each of the UAVs establishes a unified discrete progression form of the extended state in the prediction time domain based on the discrete prediction model and the discrete recursive form of the cooperative Kalman filter, so that the extended state is updated synchronously in a fixed recursive order at each prediction step.

[0066] To improve feature prediction accuracy without significantly increasing the expanded state dimension, this application adopts a consistency mechanism of "internal simplified propagation + external prediction compensation": First, the internal cooperative filtering used for online updating and prediction of the propagation feature motion model satisfies... in The selection can be based on the task; when the internal propagation complexity of the controller needs to be reduced, the "constant position" form is preferred. In order to keep the propagation structure compact.

[0067] Secondly, an external feature predictor is configured to perform more refined dynamic predictions of the features. The external predictor preferably employs a linear Gaussian enhanced state model (e.g., a constant acceleration or damped acceleration model) to better characterize the target motion and output a feature prediction sequence in the prediction time domain. This sequence is used for subsequent geometric mapping to calculate distances and angles, and based on this, to calculate the predicted measurement covariance and field-of-view weights (i.e., and ).

[0068] Finally, to ensure that the external forecast and the internal cooperative filtering are consistent in terms of time base and estimation center, this application stipulates the following synchronization rule: after the online cooperative filtering update of S400 is completed in each control cycle, the updated result is taken. As the initial value for the external predictor (and, if necessary, use...) Initialize the covariance of the external predictor with respect to the position components (generated by the external predictor). It is used consistently within the current period; at the start of the next period, the initial values ​​of the external predictor are refreshed with new online update results. This ensures that the values ​​used for calculation within the prediction time domain are consistent with the predicted values. The feature prediction sequence and the mean / covariance propagation of the internal collaborative filtering share the same update starting point, avoiding systematic deviations between "external geometric prediction" and "internal covariance propagation".

[0069] In one specific embodiment of this application, S600 includes: S610, each of the UAVs constructs a total cost function in the prediction time domain, the total cost function including the weighted sum of squared output errors obtained from the extended state mapping, the trace weighted cumulative term calculated from the estimated covariance, and the control input regularization term; Define output mapping: And provide a reference output Output It should at least include stability-related quantities for maintaining controllability, operability, and attitude stability, preferably including attitude deviation (or equivalent small-angle error) and velocity. angular velocity and thrust change rate etc.; when tasks such as positioning / formation need to be performed, it can be done in Add positional or relative position components to the task output and in The corresponding reference trajectory or reference hold value is given in the text.

[0070] To adapt to different working stages of "observation priority / motion priority", this application allows task switching to be achieved through weight matrix settings: when emphasizing active perception, the weight of position-related task components is reduced, while the weights of stable items such as attitude, velocity, and angular velocity are maintained, so that the aircraft remains stable and executable during maneuvering search.

[0071] Construct the total cost within the prediction time domain in: The output error weighting matrix; Weights for uncertainties; The input regularization weights are used to suppress unnecessary large maneuvers and high-frequency thrust changes.

[0072] The uncertainty term uses the A-optimal index. In the extended state, the covariance is... Reconstructed according to S500 during storage and computation. And Accumulation within the prediction time domain drives trajectory selection toward configurations that offer more stable field of view coverage, smaller measurement covariance, and greater cooperative information gain.

[0073] To enhance terminal performance, this application also allows for the addition of terminal items (e.g., or This does not change the above cost semantics, but only changes the temporal weight allocation.

[0074] S620, each of the aforementioned UAVs is subject to a set of constraints, which includes extended dynamic constraints based on a unified discrete propulsion form, hard constraints on thrust and thrust change rate, and unified safety interval constraints between UAVs and between UAVs and manned aircraft. The unified safety interval constraints adopt an axis-aligned safety box form and are determined by an intrusion function. When the intrusion function is greater than zero, it indicates that the UAV has entered the safety box region.

[0075] In each prediction step The following set of constraints is applied to ensure that the predicted trajectory is physically feasible and meets safety requirements.

[0076] (1) Extended dynamic constraints in It includes motor-level discrete propulsion of S200 and filter discrete propagation of S400.

[0077] (2) Actuator hard constraints During the discrete advancement process, quaternion normalization and necessary numerical consistency processing are maintained to ensure that "predictable feasibility" and "actual feasibility" are consistent.

[0078] (3) Body state constraints When the task or platform safety envelope requires it, boundary constraints can be imposed on the machine state, written as in To extract the mapping of limited quantities (such as tilt angle, velocity amplitude, angular velocity amplitude, altitude range, etc.) from the machine's state, the boundaries are set by platform capabilities and mission airspace constraints.

[0079] (4) Unified safety interval constraints for UAV-UAV and UAV-MAV To simultaneously cover the security requirements of UAV-UAV and UAV-MAV in distributed forecasting, this application uniformly refers to "security objects" as a set. in In order to be with the first The set of neighboring UAVs that have collision avoidance requirements, where "MAV" refers to manned UAVs (always considered as dynamic obstacles).

[0080] For any security object At the prediction node Define relative position in: Predict the location for this machine; For object The predicted position is preferably given directly from the predicted trajectory broadcast; when only the current state is broadcast, it can be obtained by extrapolation at a constant speed. .

[0081] The preferred approach is to use an "axis-aligned safety box exclusion region" to express a uniform safety margin constraint: given a triaxial safety half-size. Required object At any node satisfy That is, prohibiting entry to objects The space is aligned with the central axis within a safety box. This constraint aligns with the collision avoidance semantics of "applying linear distance constraints along the three coordinate axes and treating the aircraft as a box," facilitating the use of external parameters in distributed optimal control problems. The determination is made directly after injecting neighbor / manned aircraft prediction information.

[0082] To facilitate numerical processing (hard screening or soft penalty) in sampling solutions, a safety box "intrusion" function is defined. Among them: when Indicates the trajectory at the node Entering the security box (intrusion occurs); when This indicates that the safety interval is met.

[0083] The safety box half-size is determined based on the machine body envelope and safety margin. Let the first... The envelope dimensions of the UAV in the three axes of the machine system are: object The corresponding envelope size is Then the preferred option is to select in For safety margin, it is used to cover positioning error, extrapolation error, and control hysteresis; when Furthermore, when the prediction is obtained solely through extrapolation, the safety margin should be further increased according to the upper bound of the extrapolation error to maintain a conservative safety.

[0084] In one specific embodiment of this application, S700 includes: S710, each of the UAVs maintains a nominal control sequence, which is the control input sequence in the prediction time domain obtained by left shifting after the previous control cycle update, and the nominal control sequence is used as the reference sequence of the current cycle sampling predictive controller; In each control cycle , for the UAV maintenance nominal control sequence The nominal sequence is initialized using a hot start method: the nominal sequence updated in the previous control cycle is shifted one position to the left as the current initial value, and the end term is padded (filled with zero or the steady-state thrust rate of change) to improve the stability and continuity of the online solution.

[0085] S720, each of the UAVs generates a candidate control sequence by adding sampling noise based on the nominal control sequence, and sequentially performs hard constraint limiting and dynamic feasible input boundary pre-verification induced by the thrust hard constraint and the current thrust state on the candidate control input; Set the number of samples Noise covariance (dimensions and) (consistent) and temperature parameters .in: Determine the sampling coverage; Determine the search intensity; The selectivity that determines the index weighting ( The smaller the value, the more concentrated the weights are on low-cost trajectories. To ensure online feasibility, the sampling parameters are set according to the following rules: number of samples. Based on airborne computing power and control cycle Selection to ensure completion within a single cycle Subdiscrete propagation and cost calculation; noise covariance Preferably, a diagonal matrix is ​​chosen, where the diagonal elements are proportional to the magnitude of the thrust rate constraint. For example, let the first... Each channel satisfies To ensure that the sampling disturbance mainly falls within the feasible region; temperature parameter Based on the normalization setting of the cost magnitude, the index weights are not too concentrated or too uniform in value. It is preferable to adjust them in representative scenarios through offline simulation so that the closed-loop cost and the safety constraint satisfaction rate can reach the set threshold at the same time.

[0086] For each sampling trajectory In each prediction step Generate perturbations and form candidate inputs: First of all Apply input hard constraint limiting: To ensure the thrust state remains feasible throughout the prediction time domain, a "next thrust pre-verification" is introduced after each input generation step. This is based on the current thrust state. Calculate the predicted thrust in one step: Define the upper and lower bounds of the dynamic feasible input induced by the thrust hard constraint on a component-by-component basis: And perform pre-verification feasibility: Among them: the above , and Each element should be understood in terms of its components.

[0087] S730, each of the UAVs takes the current extended state as the initial value and performs forward prediction on each sampled trajectory based on the discrete prediction model. During the prediction process, the total cost is accumulated based on the objective function and the constraint set. For each sampling trajectory In the current extended state As initial value, execute Rolling forecast: (1) Motor-stage dynamics propulsion and thrust feasibility From the discrete motor-stage dynamics of the S200 propulsion system and its thrust state, we obtain... And perform hard constraint feasibility on the thrust state. (2) Calculation of the mass measured by the machine Based on predicted pose and external feature prediction (e.g.) (or its positional components), calculate the relative geometry of the machine and obtain the machine's... and .

[0088] (3) Unified safety interval determination: For each safety object , by external parameters get ,calculate When using a "hard constraint elimination" strategy, as long as any object or node satisfies... Then immediately set the total cost of the sampled trajectory to the maximum value and terminate subsequent predictions to reduce the amount of computation and strictly maintain the semantics of "no entry into the safety box".

[0089] When using a "soft constraint penalty" strategy, the intrusion quantity is transformed into a penalty term. in: The weight for security penalties should be large enough to ensure that security semantics take precedence.

[0090] After completing the recursion, for the th Total cost of each sampling trajectory S740, each of the UAVs calculates the exponential weight of all sampled trajectories. The exponential weight is calculated using an exponential function based on the difference between the total cost of each sampled trajectory and the minimum total cost. The nominal control sequence is then weighted and corrected according to the exponential weight. The first control quantity of the corrected sequence is output, and the corrected nominal control sequence is used as the initial maintenance value for the next control cycle.

[0091] Calculate the minimum cost And define exponential weights and normalized weights: Then, a weighted correction is applied to the nominal control sequence: for each prediction step renew And for the updated Re-implementing the limit feasibility This ensures that the final output control strictly meets the motor-level input constraints. The first control variable is output and executed in the current control cycle. After the control cycle ends, the updated nominal sequence is shifted left and the last term is added as the initial value for the hot start of the next control cycle, thus forming an online closed-loop operation process of sampling solution, weighted update, execution of first control and rolling shift.

[0092] In one specific embodiment of this application, S800 includes: S810, in each control cycle, each UAV uses the current extended state as the initial value, predicts its own state sequence in the future time domain based on the discrete prediction model, and updates the target state estimate and estimated covariance based on the multi-platform measurements obtained in the current cycle using the cooperative Kalman filter. S820: The updated target state estimate and estimated covariance are embedded into the extended state. The control quantity for the current cycle is solved and executed based on the sampled predictive controller. The updated extended state and the left-shifted nominal control sequence are used as the initial values ​​for the next control cycle, forming a closed loop of state perception, cooperative estimation and predictive control.

[0093] In each control cycle within, no. The UAV executes the following cyclical sequence: First, it reads the navigation estimation data obtained from the local machine. Read / estimate thrust status This forms the initial value of the extended state. ; Obtain local measurement And calculate the current according to S300. and Then, communication interaction is performed: receiving status information broadcast by the MAV, receiving measurement-related packets and prediction measurement-related packets broadcast by other UAVs; simultaneously receiving the predicted position information of neighboring UAVs used for collision avoidance determination, preferably the predicted trajectory of the neighboring UAV in the prediction time domain. Or extrapolable equivalent information, used to construct the S600 unified safety interval constraint determination. Once the information is ready, the measurements from multiple platforms are stitched together according to S400, and online collaborative estimation updates are completed to obtain... and Then update the external feature prediction. And construct the local machine in the prediction time domain The system broadcasts its own predicted measurement packets (and optional predicted trajectories) to the outside world, so that teammates can construct extrinsic parameters in their MPPI solutions. And complete the safety constraint determination. Finally, in the external parameters Once ready, use S700 to complete sampling, prediction recursion, unified safety interval determination, and exponential weighted update, output the results, and execute the program. After the control cycle ends, the nominal control sequence and cache information are shifted and updated, and the next control cycle is entered. This step is repeated until the task ends.

[0094] It is worth noting that the terms "first" and "second" in this application are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0095] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of this application and should not be construed as limiting the specific implementation of this application to these descriptions. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of this application, and all such modifications or substitutions should be considered within the scope of protection of this application.

Claims

1. A method for minimizing uncertainties in manned-heterogeneous unmanned aerial vehicle (UAV) cooperative active perception, characterized in that, include: S100, Construct a collaborative active perception system, which consists of a manned aircraft and at least one unmanned aircraft, wherein the manned aircraft periodically broadcasts its own status to the unmanned aircraft; S200: For each of the aforementioned UAVs, a continuous motor-level dynamic model is established and discretized based on the sampling period to construct a discrete prediction model; S300, for each of the aforementioned UAVs, based on the UAV pose and target feature state predicted by the discrete prediction model, construct the measurement noise covariance and continuous field of view weights. S400, each of the UAVs performs cooperative Kalman filtering based on the measurement noise covariance and the continuous field of view weight, splicing and weighting the multi-platform measurements including the manned UAV measurements, and propagating the target state estimate and estimated covariance online. S500, each of the UAVs estimates the extended state by combining the estimated covariance with the target state estimate, and establishes a unified discrete propagation form of the extended state in the prediction time domain; S600, each of the UAVs constructs a predictive control objective function and a set of constraints containing the estimated covariance trace based on the extended state and the estimated covariance trace; S700, each of the UAVs uses a sampled predictive controller to solve the problem online based on the discrete prediction model, the objective function, and the constraint set, and outputs the first control quantity of the current control cycle; S800, each of the UAVs uses the extended state as the initial value, predicts its own state using the discrete prediction model, updates the target estimate using the cooperative Kalman filter, and solves the control quantity using the sampled predictive controller. In each control cycle, a closed loop of state perception, cooperative estimation and predictive control is formed, and the extended state and nominal control sequence are updated on a rolling basis.

2. The method for minimizing uncertainties in manned-heterogeneous unmanned aerial vehicle cooperative active perception according to claim 1, characterized in that, S100 includes: S110, defines a manned aircraft periodically broadcasting its own navigation status, the navigation status including at least position, speed and attitude information; S120 defines that each UAV operates independently in a distributed communication framework for estimation and control, and the UAVs share measurement information and predicted trajectories through communication, and each UAV receives navigation status broadcast by the manned aircraft; S130, construct a collaborative active perception system for the manned aircraft and various unmanned aircraft.

3. The method for minimizing uncertainties in manned-heterogeneous unmanned aerial vehicle cooperative active perception according to claim 1, characterized in that, S200 includes: S210, for the i-th UAV, define the motor-level state and control input, and set hard constraints on the thrust vector and thrust rate of change; wherein, the motor-level state includes the airframe state and thrust vector, and the control input is the thrust rate of change; the airframe state includes position, attitude quaternion, linear velocity and angular velocity; S220, For the i-th UAV, establish a continuous motor-level dynamics model that includes rigid body dynamics and thrust state dynamics, wherein rigid body dynamics describes the changes in the state of the UAV, and thrust state dynamics describes the changes in the thrust vector with the control input; S230, for the i-th UAV, based on the rotor action point position, thrust direction unit vector, rotation direction sign and anti-torque proportional coefficient, construct the resultant force distribution matrix and resultant torque distribution matrix column by column to uniformly express the continuous motor-level dynamics model of UAVs with different configurations; S240, the continuous motor-level dynamic model is discretized based on the sampling period. In each prediction step, hard constraint limiting is applied to the control input. The dynamic feasible input boundary is calculated based on the thrust hard constraint and the current thrust state to ensure that the thrust state after one integration satisfies the thrust hard constraint. At the same time, the attitude quaternion is normalized to construct a discrete prediction model.

4. The method for minimizing uncertainties in manned-heterogeneous unmanned aerial vehicle cooperative active perception according to claim 1, characterized in that, The S300 includes: S310, For the i-th UAV, based on the UAV pose predicted by the discrete prediction model and the sensor fixed external parameters, calculate the sensor pose in the world coordinate system, extract the sensor principal axis direction, and combine the position in the target feature state to calculate the relative distance and line of sight between the sensor and the target. S320, construct a measurement noise covariance ellipsoid model coupled with the relative distance and the line of sight direction, so that the measurement uncertainty is anisotropic along the line of sight axis and the transverse tangent plane, and introduce an optimal observation distance mechanism so that the measurement noise covariance increases as the observation distance deviates from the optimal observation distance; S330, based on the cosine of the angle between the sensor's half-angle of the field of view and the deviation of the line of sight from the sensor's main axis, the continuous field of view weight is constructed using the Sigmoid function, so that the continuous field of view weight smoothly approaches one as the target enters the field of view and smoothly approaches zero as the target leaves the field of view.

5. The method for minimizing uncertainties in manned-heterogeneous unmanned aerial vehicle cooperative active perception according to claim 1, characterized in that, The S400 includes: S410, each of the UAVs splices together the measurements, observation matrices, measurement noise covariance and continuous field of view weights of each observation source in a block diagonal form, and splices together the measurements of the manned aircraft as optional observation sources to form a multi-platform spliced ​​measurement. S420, each of the UAVs performs a reliability determination on the measurements of the manned aircraft, calculates a normalized squared innovation index based on the current target state estimate, and determines an anomaly and performs rapid isolation when the normalized squared innovation index exceeds a preset threshold; otherwise, the measurement covariance of the manned aircraft is equivalently amplified according to a preset reliability weight to obtain an equivalent measurement covariance. S430, each of the UAVs uses a time-continuous Kalman filter equation to propagate the target state mean and estimated covariance, wherein the Kalman gain is calculated from the estimated covariance, the observation matrix and the equivalent measurement covariance, and is updated online in steps of the control period through discretization.

6. The method for minimizing uncertainties in manned-heterogeneous unmanned aerial vehicle cooperative active perception according to claim 3, characterized in that, The S500 includes: S510, stack the lower triangular elements of the estimated covariance matrix in a fixed order to form a minimum vector representation, and define a reconstruction mapping that reconstructs the estimated covariance matrix from the minimum vector representation; S520, each of the UAVs will combine the motor-level state, the target state estimate, and the minimum vector representation to form an extended state, wherein the motor-level state includes at least the body state and the thrust vector; S530, each of the UAVs establishes a unified discrete progression form of the extended state in the prediction time domain based on the discrete prediction model and the discrete recursive form of the cooperative Kalman filter, so that the extended state is updated synchronously in a fixed recursive order at each prediction step.

7. The method for minimizing uncertainties in manned-heterogeneous unmanned aerial vehicle cooperative active perception according to claim 6, characterized in that, The S600 includes: S610, each of the UAVs constructs a total cost function in the prediction time domain, the total cost function including the weighted sum of squared output errors obtained from the extended state mapping, the trace weighted cumulative term calculated from the estimated covariance, and the control input regularization term; S620, each of the aforementioned UAVs is subject to a set of constraints, which includes extended dynamic constraints based on a unified discrete propulsion form, hard constraints on thrust and thrust change rate, and unified safety interval constraints between UAVs and between UAVs and manned aircraft. The unified safety interval constraints adopt an axis-aligned safety box form and are determined by an intrusion function. When the intrusion function is greater than zero, it indicates that the UAV has entered the safety box region.

8. The method for minimizing uncertainties in manned-heterogeneous unmanned aerial vehicle cooperative active perception according to claim 6, characterized in that, The S700 includes: S710, each of the UAVs maintains a nominal control sequence, which is the control input sequence in the prediction time domain obtained by left shifting after the previous control cycle update, and the nominal control sequence is used as the reference sequence of the current cycle sampling predictive controller; S720, each of the UAVs generates a candidate control sequence by adding sampling noise based on the nominal control sequence, and sequentially performs hard constraint limiting and dynamic feasible input boundary pre-verification induced by the thrust hard constraint and the current thrust state on the candidate control input; S730, each of the UAVs takes the current extended state as the initial value and performs forward prediction on each sampled trajectory based on the discrete prediction model. During the prediction process, the total cost is accumulated based on the objective function and the constraint set. S740, each of the UAVs calculates the exponential weight of all sampled trajectories. The exponential weight is calculated using an exponential function based on the difference between the total cost of each sampled trajectory and the minimum total cost. The nominal control sequence is then weighted and corrected according to the exponential weight. The first control quantity of the corrected sequence is output, and the corrected nominal control sequence is used as the initial maintenance value for the next control cycle.

9. The method for minimizing uncertainties in manned-heterogeneous unmanned aerial vehicle cooperative active perception according to claim 8, characterized in that, The S800 includes: S810, in each control cycle, each UAV uses the current extended state as the initial value, predicts its own state sequence in the future time domain based on the discrete prediction model, and updates the target state estimate and estimated covariance based on the multi-platform measurements obtained in the current cycle using the cooperative Kalman filter. S820: The updated target state estimate and estimated covariance are embedded into the extended state. The control quantity for the current cycle is solved and executed based on the sampled predictive controller. The updated extended state and the left-shifted nominal control sequence are used as the initial values ​​for the next control cycle, forming a closed loop of state perception, cooperative estimation and predictive control.

10. The method for minimizing uncertainties in manned-heterogeneous unmanned aerial vehicle cooperative active perception according to claim 8, characterized in that, The estimated covariance trace, as an uncertainty measure, has its cumulative term embedded in the predictive control objective function. Through the evaluation and weighted update of candidate trajectories by the sampled predictive controller, each UAV is driven to select a flight trajectory that can reduce the estimated covariance trace, under the premise of satisfying the continuous motor-level dynamics model and the constraint set, so as to achieve the goal of minimizing the uncertainty of collaborative active perception.

Images