Heavy-load unmanned aerial vehicle cooperative control method and system for multi-point material delivery

Abstract

The application discloses a heavy-load unmanned aerial vehicle cooperative control method and system for multi-point material delivery, relates to the technical field of unmanned aerial vehicle cooperative control, and comprises the following steps: identifying a material delivery point cooperative request, analyzing a load change trend, determining a first unmanned aerial vehicle changing from heavy load to light load and a second unmanned aerial vehicle changing from light load to heavy load; collecting unmanned aerial vehicle monitoring data sets in real time, defining a six-degree-of-freedom dynamics model of the unmanned aerial vehicle, training a bidirectional compensation control model by using dynamics simulation data, and performing bidirectional compensation optimization analysis on the first unmanned aerial vehicle and the second unmanned aerial vehicle to generate first and second compensation control parameters. The technical problems that, in the prior art, the load of a multi-point cooperative delivery scene changes dynamically and sharply, single-machine control is unstable, multi-machine task coupling is poor, and a cooperative bidirectional compensation control mechanism is lacking are solved, and the technical effects of real-time cooperative optimization of unmanned aerial vehicle control parameters, improvement of the overall flight stability of multi-machine cooperative operation, task smoothness and delivery precision are achieved.

Description

Technical Field

[0001] This invention relates to the field of collaborative control technology for unmanned aerial vehicles (UAVs), specifically to a collaborative control method and system for heavy-duty UAVs used for multi-point material delivery. Background Technology

[0002] With the increasing demand for efficient and precise multi-point material delivery, the payload capacity, endurance, and mission range of a single drone are often limited in multi-point continuous delivery or collaborative operation scenarios, making it difficult to independently complete multiple differentiated delivery tasks. Collaborating with multiple heavy-load drones to form a mission cluster has become an inevitable choice to improve overall delivery efficiency, coverage, and mission robustness. However, the dynamic changes in drone payload directly impact flight attitude, energy consumption, and stability. When performing material delivery, the drone's load suddenly decreases, while when loading materials, the load suddenly increases. This step change in load significantly alters key dynamic parameters such as the drone's center of mass and moment of inertia, introducing strong nonlinear disturbances and uncertainties. Improper control can easily lead to flight instability, decreased trajectory tracking accuracy, and even mission failure. Traditional single-drone control strategies are insufficient to effectively address the frequent and drastic load changes in collaborative tasks. Furthermore, existing control methods for drone load changes are mostly focused on single-drone adaptive control or anti-interference control, which, in multi-point collaborative delivery scenarios, do not fully consider the dynamic mutual influence between multiple drones due to mission coupling, affecting the overall stability of the cluster and the smoothness of the mission.

[0003] Therefore, current technologies face technical challenges such as unstable single-machine control due to drastic dynamic changes in load during multi-point collaborative delivery scenarios, poor coupling of multi-machine tasks, and a lack of collaborative two-way compensation control mechanisms. Summary of the Invention

[0004] This application provides a collaborative control method and system for heavy-load UAVs used for multi-point material delivery. It solves the technical problems in the prior art that lead to single-machine control instability due to drastic dynamic changes in load in multi-point collaborative delivery scenarios, poor coupling of multi-machine tasks, and lack of collaborative two-way compensation control mechanism. It achieves the technical effect of real-time collaborative optimization of UAV control parameters, improving the overall flight stability, mission smoothness, and delivery accuracy of multi-machine collaborative operations.

[0005] This application provides a collaborative control method for heavy-load unmanned aerial vehicles (UAVs) for multi-point material delivery. The method includes: identifying a collaborative request task for any material delivery point; performing load change analysis on the collaborative request task to determine a first UAV and a second UAV, wherein the first UAV is a UAV whose load changes from heavy to light, and the second UAV is a UAV whose load changes from light to heavy; real-time acquisition of a first set of monitoring datasets and a second set of monitoring datasets corresponding to the first and second UAVs; defining a six-degree-of-freedom (DOF) dynamic model for the UAV; training a bidirectional compensation control model according to the dynamic simulation dataset output by the six-DOF dynamic model; and performing bidirectional compensation optimization analysis on the first and second UAVs according to the first and second sets of monitoring datasets to obtain a first compensation control parameter and a second compensation control parameter.

[0006] In a possible implementation, the heavy-load UAV collaborative control method for multi-point material delivery further performs the following processing: the UAV six-degree-of-freedom dynamic model is obtained by fitting kinematic equations to the UAV's six-degree-of-freedom body coordinate system, translational state variables, and rotational state variables; a load coupling model, environmental disturbances, and parametric case sets are introduced to simulate the operating conditions of the UAV six-degree-of-freedom dynamic model, and the dynamic simulation dataset is recorded, including state vectors, control inputs, environmental disturbance quantities, and measurement outputs; the dynamic simulation dataset is divided into a heavy-light load simulation dataset and a light-heavy load simulation dataset based on heavy-light load labels and light-heavy load labels, and a bidirectional compensation control model is trained using the heavy-light load simulation dataset and the light-heavy load simulation dataset.

[0007] In a possible implementation, the heavy-load UAV cooperative control method for multi-point material delivery further performs the following processing: extracting a first compensation optimization objective corresponding to the heavy-light load simulation dataset and a second compensation optimization objective corresponding to the light-heavy load simulation dataset; constructing a dual-branch neural network, the dual-branch neural network including a shared feature extraction layer; extracting features from the heavy-light load simulation dataset and the light-heavy load simulation dataset according to the shared feature extraction layer; performing supervised learning on the dual-branch neural network according to the first feature extraction result and the second feature extraction result to obtain first loss data and second loss data; performing supervised learning on the first loss data and the second loss data with the first compensation optimization objective and the second compensation optimization objective until the prediction accuracy output by the dual-branch neural network reaches an accuracy threshold, thereby obtaining a bidirectional compensation control model.

[0008] In a possible implementation, the heavy-duty UAV cooperative control method for multi-point material delivery further performs the following processing: training the first loss data with the first compensation optimization objective to obtain a first branch neural network whose output prediction accuracy reaches the accuracy threshold, wherein the first compensation optimization objective is a weighted fit of attitude overshoot, energy saving rate and collision avoidance risk. The second loss data is trained with the second compensation optimization objective to obtain a second branch neural network whose output prediction accuracy reaches the accuracy threshold. The second compensation optimization objective is a weighted fit of instability risk probability, power life loss and collision avoidance risk. The first branch neural network and the second branch neural network are connected to obtain a bidirectional compensation control model.

[0009] In a possible implementation, the heavy-load UAV cooperative control method for multi-point material delivery further performs the following processing: extracting the control parameter difference matrix between the heavy-light load simulation dataset and the light-heavy load simulation dataset, wherein the control parameters of the control parameter difference matrix include attitude control parameters, power distribution parameters, cooperative collision avoidance parameters, and energy consumption optimization parameters; obtaining mixed loss data of the first loss data and the second loss data; performing consistency verification on the mixed loss data based on the control parameter difference matrix, and obtaining a bidirectional compensation control model when the consistency verification passes.

[0010] In a possible implementation, the heavy-load UAV collaborative control method for multi-point material delivery further performs the following processing: extracting a first set of feature vectors and a second set of feature vectors from the first set of monitoring datasets and the second set of monitoring datasets. The first set of feature vectors includes load ratio, attitude overshoot, energy saving rate, and timestamp. The second set of feature vectors includes load ratio, attitude angular rate, energy reserve rate, and timestamp. The bidirectional compensation control model obtains a first compensation control parameter and a second compensation control parameter based on the first set of feature vectors and the second set of feature vectors.

[0011] In a possible implementation, the heavy-duty UAV collaborative control method for multi-point material delivery also performs the following processing: determining the task type of the collaborative request task; if the task type is a handover collaborative request task, controlling the first UAV and the second UAV according to the first compensation control parameter and the second compensation control parameter.

[0012] In a possible implementation, the heavy-duty UAV collaborative control method for multi-point material delivery further performs the following processing: if the task type is a synchronous collaborative request task, optimize the first compensation control parameter and the second compensation control parameter for time synchronization error, trajectory synchronization error and parameter coupling error; control the first UAV and the second UAV according to the optimized first compensation control parameter and the second compensation control parameter.

[0013] In a possible implementation, the heavy-load UAV collaborative control method for multi-point material delivery also performs the following processing: obtaining the material load mode of the first UAV and the second UAV, and retraining the bidirectional compensation control model with the material load mode as a constraint.

[0014] This application also provides a heavy-load UAV collaborative control system for multi-point material delivery. The system includes: a load change analysis module for identifying collaborative request tasks at any material delivery point, performing load change analysis on the collaborative request tasks, and determining a first UAV and a second UAV, wherein the first UAV is a UAV whose load changes from heavy to light, and the second UAV is a UAV whose load changes from light to heavy; a monitoring data acquisition module for real-time acquisition of a first set of monitoring datasets and a second set of monitoring datasets corresponding to the first UAV and the second UAV; a compensation control model training module for defining a six-degree-of-freedom dynamic model of the UAV and training a bidirectional compensation control model according to the dynamic simulation dataset output by the six-degree-of-freedom dynamic model; and a compensation optimization analysis module for performing bidirectional compensation optimization analysis on the first UAV and the second UAV according to the first set of monitoring datasets and the second set of monitoring datasets to obtain a first compensation control parameter and a second compensation control parameter.

[0015] This application proposes a heavy-load UAV collaborative control method and system for multi-point material delivery. This method identifies collaborative requests from material delivery points, analyzes load change trends, and determines the first UAV transitioning from heavy to light load and the second UAV transitioning from light to heavy load. It collects UAV monitoring datasets in real time, defines a six-degree-of-freedom dynamic model for the UAVs, trains a bidirectional compensation control model using dynamic simulation data, and performs bidirectional compensation optimization analysis on the first and second UAVs to generate first and second compensation control parameters. This solves the technical problems in existing technologies, such as single-unit control instability due to drastic dynamic load changes in multi-point collaborative delivery scenarios, poor coupling of multi-unit tasks, and the lack of a collaborative bidirectional compensation control mechanism. It achieves real-time collaborative optimization of UAV control parameters, improving the overall flight stability, mission smoothness, and delivery accuracy of multi-unit collaborative operations. Attached Figure Description

[0016] To more clearly illustrate the technical solutions of the embodiments of this disclosure, the accompanying drawings of the embodiments of this disclosure will be briefly described below. Flowcharts are used in this application to illustrate the operations performed by the system according to the embodiments of this application. It should be understood that the preceding or following operations are not necessarily performed precisely in sequence. Instead, various steps can be processed in reverse order or simultaneously as needed. Furthermore, other operations can be added to these processes, or one or more steps can be removed from these processes.

[0017] Figure 1 This is a schematic flowchart of a heavy-duty UAV collaborative control method for multi-point material delivery provided in an embodiment of this application.

[0018] Figure 2 This is a schematic diagram of the structure of a heavy-duty UAV collaborative control system for multi-point material delivery provided in an embodiment of this application.

[0019] Figure labeling: Load change analysis module 10, monitoring data acquisition module 20, compensation control model training module 30, compensation optimization analysis module 40. Detailed Implementation

[0020] To further illustrate the technical means and effects adopted by the present invention in order to achieve the intended purpose, the following detailed description is provided in conjunction with the accompanying drawings and preferred embodiments, based on the specific implementation methods, structures, features and effects of the present invention.

[0021] This application provides a method for collaborative control of heavy-duty unmanned aerial vehicles (UAVs) for multi-point material delivery, such as... Figure 1 As shown, the method includes: Step S100: Identify a collaborative request task for any material delivery point, perform load change analysis on the collaborative request task, and determine a first UAV and a second UAV. The first UAV is a UAV whose load changes from heavy load to light load, and the second UAV is a UAV whose load changes from light load to heavy load.

[0022] Preferably, the system receives and identifies collaborative request tasks from any material delivery point, dynamically classifies drone roles based on the physical execution process of the task, where a collaborative request task refers to an operational instruction initiated by a material delivery point that requires the cooperation of multiple drones to complete, specifying constraints such as the loading and unloading location, time, and weight of the materials. Then, a load change analysis is performed on the collaborative request task, i.e., analyzing the inevitable mass transfer process that occurs during the execution of the collaborative request task. For example, if the task is to pick up goods from point A and transport them to point B for delivery, the drone performing the delivery action will experience a load change from heavy to light. If the mission is to load goods from point C and transport them to point D, the drone performing the loading action will experience a change in load from light to heavy. The drone performing the material delivery action in the mission is identified as the first drone. Its onboard load decreases instantaneously due to the release of materials, causing a sudden change in the drone's dynamic parameters such as mass, center of gravity, and moment of inertia, resulting in the drone accelerating upward and experiencing attitude oscillation. The drone performing the material receiving or loading action in the mission is identified as the second drone. Its onboard load increases instantaneously due to the receiving of materials, which also causes a step change in dynamic parameters, resulting in the drone sinking downward and becoming unstable.

[0023] Step S200: Real-time acquisition of the first set of monitoring datasets and the second set of monitoring datasets corresponding to the first UAV and the second UAV.

[0024] Preferably, the first and second sets of monitoring datasets corresponding to the first and second UAVs are collected in real time through UAV onboard sensors and flight control systems. The first and second sets of monitoring datasets mainly include flight status data, power system data, payload and mission-related data, and environmental perception data. Specifically, the flight status data includes attitude data such as pitch angle, roll angle, yaw angle and its angular rate, position trajectory data such as three-dimensional coordinates, flight speed, and acceleration, fused positioning information from GPS / Inertial Measurement Unit (IMU), and heading angle. The power system data includes motor / engine data such as the rotational speed, current, voltage, temperature, and torque of each rotor or thruster, energy data such as the remaining battery power, voltage, current, temperature, and estimated flight time, and power distribution commands for each power unit currently output by the flight control system. The payload and mission-related data include payload status data and mission timing data. For the first UAV, the payload status data includes at least the remaining payload mass / proportion, the instantaneous rate of change of payload mass, the attitude overshoot and positive vertical acceleration change caused by the sudden reduction of load after release, and the additional energy consumption generated to stabilize the attitude. For the second UAV, the payload status data includes at least the target mass of the payload receiving payload, the impact force / tension when the payload is connected, the attitude descent and negative vertical acceleration change caused by the sudden increase in load, and the torque / current suddenly increased by the power system to resist descent. The mission timing data includes at least the "time of issuance of the release command," "time of cargo ground contact confirmation," and "time of hook locking." Environmental perception data includes, but is not limited to, the relative distance, relative speed, and wind speed and direction with nearby collaborating UAVs, which are determined by visual / radar or wind vane sensors, respectively, for collision avoidance and accurate positioning.

[0025] Step S300: Define a six-degree-of-freedom dynamic model for the UAV, and train a bidirectional compensation control model according to the dynamic simulation dataset output by the six-degree-of-freedom dynamic model.

[0026] Step S300 further includes step S310, whereby the UAV six-degree-of-freedom dynamic model is obtained by fitting kinematic equations to the UAV six-degree-of-freedom body coordinate system, translational state variables, and rotational state variables; step S320, by introducing a load coupling model, environmental disturbances, and parametric working condition sets, the UAV six-degree-of-freedom dynamic model is simulated under working conditions, and the dynamic simulation dataset is recorded, including state vectors, control inputs, environmental disturbance quantities, and measurement outputs; step S330, the dynamic simulation dataset is divided into a heavy-light load simulation dataset and a light-heavy load simulation dataset based on heavy-light load labels and light-heavy load labels, and a bidirectional compensation control model is trained using the heavy-light load simulation dataset and the light-heavy load simulation dataset.

[0027] Preferably, kinematic equations are fitted to the UAV's six-degree-of-freedom body coordinate system, translational state variables, and rotational state variables. Specifically, a body coordinate system is established for the UAV itself, with its origin at the center of mass, the X-axis pointing forward, the Y-axis pointing left, and the Z-axis pointing upward, as well as a geodetic coordinate system describing its position relative to the ground. The translational state variables include the components of position (x, y, z) and velocity (u, v, w) on the three axes, and the rotational state variables include the components of attitude angles (pitch angle θ, roll angle φ, yaw angle ψ) and angular velocities (p, q, r) on the three axes. Based on the Newton-Euler equations, a set of differential equations is established to describe the relationship between the six position / attitude and six velocity / angular velocity state variables and the gravity, engine thrust, air resistance, and torque acting on the UAV. This serves as the standard mathematical model for UAV flight dynamics. The parameters of the equation set are determined according to the UAV's physical parameters, thereby defining the UAV's six-degree-of-freedom dynamic model, which can accurately describe the UAV's motion in three-dimensional space.

[0028] Preferably, the load coupling model is used to describe the impact of load on the overall dynamic characteristics of the UAV, including the sudden change in the total mass of the UAV caused by load loading and unloading, the change in the overall center of gravity coordinates of the UAV caused by load position shift, the inertia of the UAV rotating around each axis due to changes in load distribution and weight. The load coupling model takes the mass, position, and connection method of the load as input and dynamically calculates its impact on the mass and gravity terms in the basic dynamic equations. Environmental disturbances refer to the interference factors added in the simulation, such as different wind directions and speeds, atmospheric turbulence, etc. The parametric case set is a structured set of historical scene parameters used to traverse various flight conditions, which may include flight states such as hovering, forward flight, climb, and turning, load parameters such as load weight and center offset, and flight stages corresponding to load changes, etc. Then, load coupling model, environmental disturbance and parametric case set are introduced to simulate the working conditions of the UAV six-degree-of-freedom dynamic model. In each simulation, the state vector, that is, the value of all translational and rotational state variables at a certain moment, is recorded according to the time step, the control input, that is, the virtual control command applied to the model at that moment, the environmental disturbance, that is, the simulated wind speed, wind direction, etc., at that moment, and the measurement output, that is, the state data monitored by the simulated sensor, are formed into a dynamic simulation dataset.

[0029] Preferably, the simulation process is determined to be either deployment or loading, and a heavy-light load label and a light-heavy load label are assigned to the simulation operation. Based on the heavy-light load label and the light-heavy load label, the dynamic simulation dataset is divided into a heavy-light load simulation dataset and a light-heavy load simulation dataset, corresponding to the determined first UAV and second UAV, respectively. Then, the heavy-light load simulation dataset and the light-heavy load simulation dataset are used as training samples to train a two-branch neural network to obtain a bidirectional compensation control model. The heavy-light load simulation dataset is used to train the neural network branch of the model to handle the unloading situation, and the light-heavy load simulation dataset is used to train the neural network branch of the model to handle the loading situation. By learning from the data generated by the high-fidelity physical model, the bidirectional compensation control model can determine the optimal compensation control parameters based on monitoring data under various complex working conditions.

[0030] Further, step S330 also includes step S331, extracting the first compensation optimization objective corresponding to the heavy-light load simulation dataset and the second compensation optimization objective of the light-heavy load simulation dataset respectively; step S332, constructing a dual-branch neural network, the dual-branch neural network including a shared feature extraction layer; step S333, performing feature extraction on the heavy-light load simulation dataset and the light-heavy load simulation dataset according to the shared feature extraction layer, and performing supervised learning on the dual-branch neural network according to the first feature extraction result and the second feature extraction result to obtain first loss data and second loss data; step S334, performing supervised learning on the first loss data and the second loss data with the first compensation optimization objective and the second compensation optimization objective until the prediction accuracy output by the dual-branch neural network reaches the accuracy threshold, thereby obtaining a bidirectional compensation control model.

[0031] Preferably, the heavy-light load simulation dataset corresponds to the unloading machine / first UAV, and the corresponding first compensation optimization objective is extracted, which may include a multi-objective weighted average of minimizing attitude overshoot, maximizing energy saving rate, and minimizing collision avoidance risk; the light-heavy load simulation dataset corresponds to the loading machine / second UAV, and the corresponding second compensation optimization objective is extracted, which may include a multi-objective weighted average of minimizing instability risk probability, minimizing power system life loss, and minimizing collision avoidance risk. A dual-branch neural network with two independent output heads is constructed. The first branch neural network is responsible for learning and generating unloading compensation parameters, and the second branch neural network is responsible for learning and generating loading compensation parameters. The dual-branch neural network includes a shared feature extraction layer, that is, the first branch neural network and the second branch neural network share the same set of bottom neural network layers at the network front end, which is used to extract high-order features related to flight state and load changes from the original monitoring data, thereby significantly reducing the number of model parameters and improving efficiency.

[0032] Preferably, the heavy-light load simulation dataset and the light-heavy load simulation dataset are respectively input into a shared feature extraction layer for feature extraction, obtaining a first feature extraction result and a second feature extraction result. Then, supervised learning is performed on the dual-branch neural network based on the first and second feature extraction results. Specifically, the first feature extraction result is input into the first branch for learning to obtain a first prediction parameter, which is compared with a first compensation optimization objective, and the deviation is calculated as the first loss data. Similarly, the second feature extraction result is input into the second branch for learning to obtain a second prediction parameter, which is compared with a second compensation optimization objective, and the deviation is calculated as the second loss data, used to quantify the error level of the model's current prediction. Then, supervised learning is performed on the first and second loss data using the first and second compensation optimization objectives, i.e., using the first and second loss data respectively, gradient descent optimization is used to inversely adjust the parameters of the first branch neural network, the second branch neural network, and the shared feature layer, reducing the first and second loss data. Iterative optimization is performed so that the prediction compensation parameters output by the two branch neural networks increasingly conform to the first and second compensation optimization objectives. When the model predicts compensation parameters on simulation test data that enable the actual performance indicators of the UAV to reach the preset accuracy threshold, training stops, and a neural network with fixed parameters is obtained, which is the final deployable bidirectional compensation control model. It can judge complex working conditions and give the most suitable compensation control parameters based on real-time monitoring data.

[0033] Furthermore, step S334 also includes training the first loss data with the first compensation optimization objective to obtain a first branch neural network whose output prediction accuracy reaches an accuracy threshold, wherein the first compensation optimization objective is a weighted fit of attitude overshoot, energy saving rate and collision avoidance risk; training the second loss data with the second compensation optimization objective to obtain a second branch neural network whose output prediction accuracy reaches an accuracy threshold, wherein the second compensation optimization objective is a weighted fit of instability risk probability, power life loss and collision avoidance risk; and connecting the first branch neural network and the second branch neural network to obtain a bidirectional compensation control model.

[0034] Preferably, the first compensation optimization objective is a weighted fit of attitude overshoot, energy saving rate, and collision avoidance risk. Attitude overshoot is the maximum oscillation amplitude of the drone's attitude relative to the target stable value after cargo delivery. Energy saving rate is the ratio of the extra energy consumed during attitude recovery and stabilization to the baseline energy consumption. Collision avoidance risk is the probability estimate of collision with surrounding obstacles or other drones during dynamic adjustment. Different weights are assigned to these parameters according to task priority, and the weighted fit determines the first total loss function. The heavy-light load simulation dataset is processed through a shared feature extraction layer, input into the first branch neural network, and outputs the first prediction parameters. These parameters are then applied to the drone in the simulation environment to calculate the actual attitude overshoot, energy saving rate, and collision avoidance risk after simulated flight. This calculates the first loss data for the simulation. Backpropagation updates the parameters of the first branch neural network and the shared feature layer, iterating repeatedly until the predictions of the first branch neural network enable the average performance of the simulated drone on the test set to reach a preset accuracy threshold, such as attitude overshoot less than 5 degrees and energy saving rate higher than 90%.

[0035] Preferably, the second compensation optimization objective is a weighted fit of the instability risk probability, power life loss, and collision avoidance risk. The instability risk probability is the probability that the UAV will lose attitude control due to sudden changes in load and center of gravity when receiving a heavy object. Power life loss is the fatigue accumulation and heat loss caused by the motor / engine's instantaneous output of high torque to resist descent. Collision avoidance risk is used to ensure the safety of collaborative operations. Different weights are assigned to these factors according to task priority. The weighted fit determines the second total loss function. Similarly, a light-heavy load simulation dataset is used for simulation prediction, outputting the second loss data. The parameters of the second branch neural network and the shared feature layer are updated through backpropagation. This process is iterated until the prediction of the second branch neural network enables the average performance of the simulated UAV on the test set to reach a preset accuracy threshold. Finally, the first and second branch neural networks are connected, and a shared feature extraction layer is configured to obtain a bidirectional compensation control model. When the current UAV is identified as the first UAV, the real-time monitoring data is processed by the shared feature layer and routed to the first branch neural network to generate the first compensation control parameters. When the current UAV is identified as the second UAV, the data is routed to the second branch neural network to generate the second compensation control parameters.

[0036] Furthermore, step S334 also includes extracting the control parameter difference matrix between the heavy-light load simulation dataset and the light-heavy load simulation dataset. The control parameters in the control parameter difference matrix include attitude control parameters, power distribution parameters, cooperative collision avoidance parameters, and energy consumption optimization parameters. It also includes obtaining mixed loss data of the first loss data and the second loss data; performing consistency verification on the mixed loss data based on the control parameter difference matrix; and obtaining a bidirectional compensation control model when the consistency verification passes.

[0037] Preferably, control parameter vectors are extracted from the heavy-load and light-load simulation datasets under cooperative scenarios, respectively. These control parameters include attitude control parameters, such as desired attitude angles, angular velocities, or corresponding PID controller gains; power distribution parameters, such as thrust / torque commands for each motor or control surface; cooperative collision avoidance parameters, such as additional speed or heading corrections to maintain formation or avoid collisions; and energy optimization parameters, such as motor operating point offsets related to energy-saving strategies. The difference between two control parameter vectors in a large number of paired cooperative scenario samples is repeatedly calculated, and their covariance matrix is ​​calculated to generate a control parameter difference matrix, characterizing the control parameters of the two types of UAVs during typical cooperative interactions. The normal range or expected pattern of the difference is defined; the first loss data and the second loss data are fused to determine the mixed loss data, which comprehensively reflects the overall performance on the two types of tasks; then the mixed loss data is validated for consistency based on the control parameter difference matrix, which is used as a regularization term of the loss function. For the current model parameters, the actual control parameter difference generated on a large number of paired collaborative samples is calculated again, and the distance or degree of difference between the actual control parameter difference and the expected / reasonable difference pattern defined by the control parameter difference matrix is ​​evaluated. When the consistency validation is passed, training stops and a bidirectional compensation control model is obtained, thereby ensuring improved practicality and reliability in real complex collaborative tasks.

[0038] In step S400, the bidirectional compensation control model performs bidirectional compensation optimization analysis on the first UAV and the second UAV according to the first set of monitoring datasets and the second set of monitoring datasets to obtain the first compensation control parameters and the second compensation control parameters.

[0039] Step S400 further includes step S410, extracting a first set of feature vectors and a second set of feature vectors from the first set of monitoring datasets and the second set of monitoring datasets. The first set of feature vectors includes load ratio, attitude overshoot, energy saving rate, and timestamp. The second set of feature vectors includes load ratio, attitude angular rate, energy reserve rate, and timestamp. Step S420, the bidirectional compensation control model obtains a first compensation control parameter and a second compensation control parameter based on the first set of feature vectors and the second set of feature vectors.

[0040] Preferably, the shared feature extraction layer within the bidirectional compensation control model extracts the first set of feature vectors and the second set of feature vectors corresponding to the first set of monitoring datasets and the second set of monitoring datasets, respectively. The first set of feature vectors includes load ratio, attitude overshoot, energy saving rate, and timestamp. The load ratio is the ratio of the current remaining load mass to the maximum load capacity, used to determine the unloading process and remaining inertia. The attitude overshoot is the maximum positive deviation between the actual attitude angle and the target attitude angle in the most recent control cycle, used to quantify the degree of instability after unloading. The energy saving rate is the ratio of the actual energy consumption to the estimated energy consumption under uncompensated control from the start of the deployment action, reflecting the energy efficiency level of the current control strategy. The timestamp is used to determine the stage of the dynamic process. The second set of feature vectors includes payload ratio, attitude angular rate, energy reserve rate, and timestamp. The payload ratio is the ratio of the currently received payload mass to the target payload mass, used to determine the loading progress and the addition of inertia. The attitude angular rate is the angular velocity of the UAV fuselage rotating around each axis. At the moment of loading, a sudden change in angular rate is a precursor to instability risk. The energy reserve rate is the ratio of the currently remaining available energy to the estimated total energy required to complete the current task, used to assess the endurance risk of the power system. The timestamp is used for timing judgment and coordination.

[0041] Preferably, the first set of feature vectors and the second set of feature vectors are input into the bidirectional compensation control model to perform bidirectional compensation optimization analysis on the first UAV and the second UAV. Specifically, the first set of feature vectors of the first UAV is routed to the first branch neural network, and the second set of feature vectors of the second UAV is routed to the second branch neural network. After receiving the data, the first branch neural network performs forward propagation calculation and outputs optimized first compensation control parameters to deal with unloading conditions. These parameters may include target angular velocity commands to suppress attitude overshoot, motor thrust distribution schemes optimized for energy saving, and trajectory fine-tuning considering cooperative collision avoidance. Similarly, the second branch neural network outputs second compensation control parameters to deal with loading conditions. These parameters may include target torque commands to resist sudden changes in attitude angular rate, smoothed motor torque commands to protect power life, and hovering position holding parameters to ensure safe loading.

[0042] Furthermore, step S400 also includes determining the task type of the collaborative request task; if the task type is a handover collaborative request task, controlling the first UAV and the second UAV according to the first compensation control parameter and the second compensation control parameter.

[0043] Preferably, the collaborative request task instruction is parsed, and the collaborative operation mode of the task is determined based on the spatiotemporal logical relationship. If the task type is a handover collaborative request task, that is, the operation actions of two or more UAVs are closely connected in time and space and there is a clear material or responsibility transfer relationship, such as the relay transfer of materials, then the first compensation control parameter and the second compensation control parameter are directly sent to the flight control systems of the first UAV and the second UAV, and combined with the basic flight states such as position loop, attitude loop PID control, etc., to generate the final motor instruction, which drives the UAV body to complete the precise delivery and stable reception actions. At the same time, it ensures that the unloader smoothly puts down the cargo to the greatest extent and avoids violent shaking that could hit the receiver or cargo; the loader stably catches the cargo and avoids instability due to impact or collision with the unloader, thereby safely, efficiently and smoothly realizing the transfer of heavy-load materials between UAVs.

[0044] Furthermore, step S400 also includes, if the task type is a synchronous collaborative request task, optimizing the first compensation control parameter and the second compensation control parameter for time synchronization error, trajectory synchronization error and parameter coupling error; controlling the first UAV and the second UAV according to the optimized first compensation control parameter and the second compensation control parameter.

[0045] Preferably, if the task type is a synchronous collaborative request task, that is, an operation in which multiple drones perform actions in parallel and coordinate with each other to achieve the same goal within the same time period, such as multiple drones simultaneously dropping supplies to different locations or multiple drones forming a formation for coordinated transportation, then the first and second compensation control parameters are optimized for time synchronization error, trajectory synchronization error, and parameter coupling error. Specifically, the expected time point for each drone to perform actions based on its own parameters is calculated and compared with the time point required by the global command to obtain the time synchronization error. The time-related parts in the compensation control parameters of each drone are fine-tuned so that the expected key action time points of all drones converge to be consistent, ensuring simultaneity. Based on the predicted trajectory of each drone, the trajectory synchronization error between them is calculated. Errors such as relative position, speed, and heading are addressed by adding additional trajectory corrections to the compensation control parameters of each UAV to minimize the overall formation error. This ensures that even when individuals perform dynamic compensation, the group maintains the preset formation or relative motion relationship, ensuring consistency. A dynamic coupling model describing the task coupling between multiple UAVs is established to analyze the current combination of the first and second compensation control parameters, predict group-level instability or inefficiency, and then jointly iteratively optimize the compensation control parameters of all UAVs with the goal of optimizing the overall group performance, ensuring integrity. The optimized first and second compensation control parameters are obtained and sent to the first and second UAVs for execution, thereby achieving the best balance between individual stable compensation and group cooperative constraints.

[0046] Furthermore, step S400 also includes obtaining the material load mode of the first UAV and the second UAV, and retraining the bidirectional compensation control model with the material load mode as a constraint.

[0047] Preferably, the material loading method refers to the specific mechanism of physical connection and force transmission between the UAV and the load, which directly affects the dynamic characteristics during load abrupt changes. This may include: suspended loading, where materials are suspended below the UAV via ropes, straps, or electromagnetic chucks; in-cabin / platform loading, where materials are placed in the cargo bay inside the UAV or on a platform on top; and robotic arm grasping loading, where materials are grasped and released by a robotic arm. The specific material loading method used by each UAV in the current mission is obtained through mission configuration files, UAV model databases, or visual recognition. Then, in the six-degree-of-freedom dynamic model used to generate training data, the models corresponding to the identified material loading methods, such as the differential equations for suspended swaying, the sliding friction model for in-cabin cargo, and the robotic arm dynamic model, are introduced as new precise constraints. Simulations generate heavy-light load and light-heavy load simulation datasets reflecting the characteristics of the loading method. Supervised retraining is then performed starting with the bidirectional compensation control model, fine-tuning the parameters of the branch neural network closely related to the output control strategy, ultimately obtaining an updated bidirectional compensation control model. This model can cope with specific disturbances such as swaying, generating a control strategy highly compatible with the current physical hardware configuration, improving the accuracy and safety of UAV collaborative control.

[0048] In the above text, refer to Figure 1 This paper describes in detail a heavy-load UAV cooperative control method for multi-point material delivery according to embodiments of the present invention. Next, reference will be made to... Figure 2 A heavy-duty unmanned aerial vehicle (UAV) collaborative control system for multi-point material delivery according to an embodiment of the present invention is described.

[0049] The heavy-load UAV collaborative control system for multi-point material delivery according to embodiments of the present invention addresses the technical problems in existing technologies, such as single-unit control instability caused by drastic dynamic changes in load during multi-point collaborative delivery scenarios, poor coupling of multi-unit tasks, and lack of a collaborative two-way compensation control mechanism. It achieves real-time collaborative optimization of UAV control parameters, improving the overall flight stability, mission smoothness, and delivery accuracy of multi-unit collaborative operations. Figure 2 As shown, the heavy-duty UAV collaborative control system for multi-point material delivery includes: a load change analysis module 10, a monitoring data acquisition module 20, a compensation control model training module 30, and a compensation optimization analysis module 40.

[0050] The load change analysis module 10 is used to identify the collaborative request task for any material delivery point, perform load change analysis on the collaborative request task, and determine the first UAV and the second UAV. The first UAV is the UAV whose load changes from heavy load to light load, and the second UAV is the UAV whose load changes from light load to heavy load. The monitoring data acquisition module 20 is used to collect the first set of monitoring datasets and the second set of monitoring datasets corresponding to the first UAV and the second UAV in real time. The compensation control model training module 30 is used to define a six-degree-of-freedom dynamic model of the UAV and train a bidirectional compensation control model according to the dynamic simulation dataset output by the six-degree-of-freedom dynamic model. The compensation optimization analysis module 40 is used to perform bidirectional compensation optimization analysis on the first UAV and the second UAV according to the first set of monitoring datasets and the second set of monitoring datasets by the bidirectional compensation control model to obtain the first compensation control parameter and the second compensation control parameter.

[0051] The specific configuration of the compensation control model training module 30 will be described in detail below. The compensation control model training module 30 further includes: the UAV six-degree-of-freedom dynamic model is obtained by fitting kinematic equations to the UAV's six-degree-of-freedom body coordinate system, translational state variables, and rotational state variables; a load coupling model, environmental disturbances, and parametric case sets are introduced to simulate the operating conditions of the UAV six-degree-of-freedom dynamic model, recording the dynamic simulation dataset, including state vectors, control inputs, environmental disturbance quantities, and measurement outputs; the dynamic simulation dataset is divided into a heavy-light load simulation dataset and a light-heavy load simulation dataset based on heavy-light load and light-heavy load labels, and a bidirectional compensation control model is trained using the heavy-light load simulation dataset and the light-heavy load simulation dataset.

[0052] The specific configuration of the compensation control model training module 30 will be described in detail below. The compensation control model training module 30 further includes: extracting a first compensation optimization objective corresponding to the heavy-light load simulation dataset and a second compensation optimization objective corresponding to the light-heavy load simulation dataset; constructing a dual-branch neural network, the dual-branch neural network including a shared feature extraction layer; performing feature extraction on the heavy-light load simulation dataset and the light-heavy load simulation dataset according to the shared feature extraction layer; performing supervised learning on the dual-branch neural network according to the first feature extraction result and the second feature extraction result to obtain first loss data and second loss data; performing supervised learning on the first loss data and the second loss data using the first compensation optimization objective and the second compensation optimization objective until the prediction accuracy output by the dual-branch neural network reaches an accuracy threshold, thereby obtaining a bidirectional compensation control model.

[0053] The specific configuration of the compensation control model training module 30 will be described in detail below. The compensation control model training module 30 further includes: a first branch neural network trained on the first loss data using the first compensation optimization objective to obtain an output prediction accuracy reaching an accuracy threshold, wherein the first compensation optimization objective is a weighted fit of attitude overshoot, energy saving rate, and collision avoidance risk; a second branch neural network trained on the second loss data using the second compensation optimization objective to obtain an output prediction accuracy reaching an accuracy threshold, wherein the second compensation optimization objective is a weighted fit of instability risk probability, power life loss, and collision avoidance risk; and a bidirectional compensation control model obtained by connecting the first branch neural network and the second branch neural network.

[0054] The specific configuration of the compensation control model training module 30 will be described in detail below. The compensation control model training module 30 further includes: extracting the control parameter difference matrix between the heavy-light load simulation dataset and the light-heavy load simulation dataset, wherein the control parameters in the control parameter difference matrix include attitude control parameters, power allocation parameters, cooperative collision avoidance parameters, and energy consumption optimization parameters; obtaining mixed loss data of the first loss data and the second loss data; performing consistency verification on the mixed loss data based on the control parameter difference matrix; and obtaining a bidirectional compensation control model when the consistency verification passes.

[0055] The specific configuration of the compensation optimization analysis module 40 will be described in detail below. The compensation optimization analysis module 40 further includes: extracting a first set of feature vectors and a second set of feature vectors from the first set of monitoring datasets and the second set of monitoring datasets. The first set of feature vectors includes load ratio, attitude overshoot, energy saving rate, and timestamp; the second set of feature vectors includes load ratio, attitude angular rate, energy reserve rate, and timestamp. The bidirectional compensation control model obtains a first compensation control parameter and a second compensation control parameter based on the first set of feature vectors and the second set of feature vectors.

[0056] The specific configuration of the compensation optimization analysis module 40 will be described in detail below. The compensation optimization analysis module 40 further includes: determining the task type of the collaborative request task; if the task type is a handover collaborative request task, controlling the first UAV and the second UAV according to the first compensation control parameter and the second compensation control parameter.

[0057] The specific configuration of the compensation optimization analysis module 40 will be described in detail below. The compensation optimization analysis module 40 further includes: if the task type is a synchronous collaborative request task, optimizing the first compensation control parameters and the second compensation control parameters for time synchronization error, trajectory synchronization error, and parameter coupling error; and controlling the first and second UAVs according to the optimized first and second compensation control parameters.

[0058] The specific configuration of the compensation optimization analysis module 40 will be described in detail below. The compensation optimization analysis module 40 further includes: acquiring the material load modes of the first UAV and the second UAV, and retraining the bidirectional compensation control model using the material load modes as constraints.

[0059] The heavy-duty UAV collaborative control system for multi-point material delivery provided in this embodiment of the invention can execute the heavy-duty UAV collaborative control method for multi-point material delivery provided in any embodiment of the invention, and has the corresponding functional modules and beneficial effects of the execution method.

[0060] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A collaborative control method for heavy-load unmanned aerial vehicles (UAVs) used for multi-point material delivery, characterized in that, The method includes: Identify any collaborative request task for material delivery point, perform load change analysis on the collaborative request task, and determine the first UAV and the second UAV. The first UAV is a UAV whose load changes from heavy load to light load, and the second UAV is a UAV whose load changes from light load to heavy load. Real-time acquisition of the first set of monitoring datasets and the second set of monitoring datasets corresponding to the first UAV and the second UAV; Define a six-degree-of-freedom dynamic model for the UAV, and train a bidirectional compensation control model according to the dynamic simulation dataset output by the six-degree-of-freedom dynamic model; The bidirectional compensation control model performs bidirectional compensation optimization analysis on the first UAV and the second UAV according to the first set of monitoring datasets and the second set of monitoring datasets to obtain the first compensation control parameters and the second compensation control parameters. The method for training a bidirectional compensation control model based on the dynamic simulation dataset output by the six-degree-of-freedom dynamic model includes: The six-degree-of-freedom dynamic model of the UAV is obtained by fitting the kinematic equations of the UAV's six-degree-of-freedom body coordinate system, translational state variables, and rotational state variables. By introducing a load coupling model, environmental disturbances, and a parametric case set, the operating conditions of the UAV's six-degree-of-freedom dynamic model are simulated, and the dynamic simulation dataset is recorded, including state vectors, control inputs, environmental disturbances, and measurement outputs. The dynamic simulation dataset is divided into a heavy-light load simulation dataset and a light-heavy load simulation dataset based on the heavy-light load label and the light-heavy load label, and a bidirectional compensation control model is trained using the heavy-light load simulation dataset and the light-heavy load simulation dataset. The method for training a bidirectional compensation control model using the heavy-light load simulation dataset and the light-heavy load simulation dataset includes: Extract the first compensation optimization objective corresponding to the heavy-light load simulation dataset and the second compensation optimization objective corresponding to the light-heavy load simulation dataset respectively; Construct a dual-branch neural network, wherein the dual-branch neural network includes a shared feature extraction layer; The shared feature extraction layer is used to extract features from the heavy-light load simulation dataset and the light-heavy load simulation dataset. The dual-branch neural network is then supervised learning based on the first feature extraction result and the second feature extraction result to obtain the first loss data and the second loss data. The first and second loss data are subjected to supervised learning using the first and second compensation optimization objectives until the prediction accuracy of the dual-branch neural network output reaches the accuracy threshold, thus obtaining a bidirectional compensation control model.

2. The heavy-load UAV collaborative control method for multi-point material delivery as described in claim 1, characterized in that, Supervised learning is performed on the first loss data and the second loss data using the first compensation optimization objective and the second compensation optimization objective, the method including: The first loss data is trained with the first compensation optimization objective to obtain a first branch neural network whose output prediction accuracy reaches the accuracy threshold. The first compensation optimization objective is a weighted fit of attitude overshoot, energy saving rate and collision avoidance risk. The second loss data is trained with the second compensation optimization objective to obtain a second branch neural network whose output prediction accuracy reaches the accuracy threshold. The second compensation optimization objective is a weighted fit of instability risk probability, power life loss and collision avoidance risk. Connecting the first branch neural network and the second branch neural network yields a bidirectional compensation control model.

3. The heavy-load UAV collaborative control method for multi-point material delivery as described in claim 1, characterized in that, The method further includes supervising the learning of the first loss data and the second loss data using the first compensation optimization objective and the second compensation optimization objective, and performing the learning on the first loss data and the second loss data. Extract the control parameter difference matrix between the heavy-light load simulation dataset and the light-heavy load simulation dataset. The control parameters in the control parameter difference matrix include attitude control parameters, power distribution parameters, cooperative collision avoidance parameters, and energy consumption optimization parameters. Obtain mixed loss data of the first loss data and the second loss data; The consistency of the mixed loss data is verified based on the control parameter difference matrix. When the consistency verification is successful, a two-way compensation control model is obtained.

4. The heavy-load UAV collaborative control method for multi-point material delivery as described in claim 1, characterized in that, The bidirectional compensation control model performs bidirectional compensation optimization analysis on the first and second UAVs according to the first set of monitoring datasets and the second set of monitoring datasets. The method includes: Extract the first set of feature vectors and the second set of feature vectors from the first set of monitoring datasets and the second set of monitoring datasets. The first set of feature vectors includes load ratio, attitude overshoot, energy saving rate and timestamp. The second set of feature vectors includes load ratio, attitude angular rate, energy reserve rate and timestamp. The bidirectional compensation control model obtains the first compensation control parameter and the second compensation control parameter based on the first set of feature vectors and the second set of feature vectors.

5. The heavy-load UAV collaborative control method for multi-point material delivery as described in claim 1, characterized in that, After identifying the first and second drones, the method further includes: Determine the task type of the collaborative request task; If the task type is a handover and collaboration request task, the first UAV and the second UAV are controlled according to the first compensation control parameter and the second compensation control parameter.

6. The heavy-load UAV collaborative control method for multi-point material delivery as described in claim 5, characterized in that, If the task type is a synchronous collaborative request task, optimize the first compensation control parameter and the second compensation control parameter for time synchronization error, trajectory synchronization error and parameter coupling error. The first and second UAVs are controlled according to the optimized first and second compensation control parameters.

7. The heavy-load UAV collaborative control method for multi-point material delivery as described in claim 1, characterized in that, After identifying the first and second drones, the method further includes: Obtain the material load mode of the first UAV and the second UAV, and retrain the bidirectional compensation control model with the material load mode as a constraint.

8. A heavy-duty UAV collaborative control system for multi-point material delivery, characterized in that, The system is used to implement the heavy-load UAV collaborative control method for multi-point material delivery as described in any one of claims 1 to 7, and the system includes: The load change analysis module is used to identify the collaborative request task of any material delivery point, perform load change analysis on the collaborative request task, and determine the first UAV and the second UAV. The first UAV is the UAV whose load changes from heavy load to light load, and the second UAV is the UAV whose load changes from light load to heavy load. The monitoring data acquisition module is used to collect the first set of monitoring datasets and the second set of monitoring datasets corresponding to the first UAV and the second UAV in real time. The compensation control model training module is used to define a six-degree-of-freedom dynamic model of the UAV and train a bidirectional compensation control model according to the dynamic simulation dataset output by the six-degree-of-freedom dynamic model. The compensation optimization analysis module is used by the bidirectional compensation control model to perform bidirectional compensation optimization analysis on the first UAV and the second UAV according to the first set of monitoring datasets and the second set of monitoring datasets, so as to obtain the first compensation control parameters and the second compensation control parameters.

Images