Quadrotor unmanned aerial vehicle cooperative transportation control method based on deep neural network
By fitting a nonlinear model predictive controller based on a deep neural network method, the problem of high computational load in the cooperative transportation control of quadrotor UAVs is solved, and efficient real-time control is achieved.
Patent Information
- Application Number
- CN202510237743.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-02-28
- Publication Date
- 2025-10-17
AI Technical Summary
Existing quadcopter UAV collaborative transport control methods require a large amount of computation in practical applications, making them difficult to deploy in real time on equipment, and the computational power requirements of nonlinear model predictive controllers are too high.
A deep neural network is used to fit the nonlinear model predictive controller. By constructing the dynamic equations of the cooperative transportation of quadrotor UAVs, a nonlinear model predictive controller is designed, and the deep neural network is trained using flight data to reduce the computational resource requirements.
This approach achieves a significant reduction in computational resource requirements and improves the real-time performance and accuracy of control while preserving the control performance of the nonlinear model predictive controller.
Smart Images
Figure CN120802971A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application belongs to the field of unmanned aerial vehicle cooperative transportation control, and particularly relates to a quadrotor unmanned aerial vehicle cooperative transportation control method based on a deep neural network. BACKGROUND
[0002] Quadrotor UAVs have been widely used in aerial photography, transportation, agriculture, and inspection due to their low cost, vertical take-off and landing, and small size. In recent years, researchers have conducted in-depth research on the cooperative transportation control of quadrotor UAVs. For example, Maza et al. (Maza I, Kondak K, Bernard M, et al. Multi-UAV Cooperation and Control for Load Transportation and Deployment [C]. 2010 The 2nd International Symposium on Unmanned Aerial Vehicles (UAVs), Reno, Nevada, USA, 2010.) designed a proportional-integral-derivative controller, but the PID controller lacks the ability to reduce load oscillation in under-damped systems. Bacelar et al. (Bacelar T, Madeiras J, Melicio R, et al. On-Board Implementation and Experimental Validation of Collaborative Transportation of Loads with Multiple UAVs [J]. Aerospace Science and Technology, 2020, 107:106284.) designed a controller combining a linear quadratic regulator and a Kalman filter, and verified the effectiveness of the controller through experiments. Ariyibi et al. (Ariyibi S O, Tekinalp O. Quaternion-Based Nonlinear Attitude Control of Quadrotor Formations Carrying A Slung Load [J]. Aerospace Science and Technology, 2020, 105: 105995.) proposed a hierarchical control system that uses a linear controller in the position control loop and a quaternion-based nonlinear controller in the attitude control loop. The performance of the controller was verified through simulation tests on rigid and flexible loads in a simulation environment.Wehbeh et al. (Wehbeh J, Rahman S, Sharf I. Distributed Model Predictive Control for UAVs Collaborative Payload Transport [C]. 2020 IEEE / RSJ International Conference on Intelligent Robots and Systems (IROS), Las Vegas, Nevada, USA, 2020.) designed a distributed collaborative transport method based on model predictive control, further considering the nonlinear characteristics of the system in the control range. Lee et al. (Lee T, Sreenath K, Kumar V, et al. Geometric Control of Cooperating Multiple Quadrotor UAVs with a Suspended Payload [C]. 2013 52nd IEEE Annual Conference on Decision and Control (CDC), Florence, ITALY, 2013.) extended the geometric controller to multiple quadrotor systems, based on the nonlinear coupled model of the multiple quadrotor suspended transport system derived by the Euler-Lagrange method, successfully solved the trajectory tracking problem of any number of flying robots suspended transport rigid load. The simulation results show that the position and attitude tracking error of the load is small, and the position and attitude tracking error of each quadrotor is also small, and the quadrotor and the load can track the expected trajectory well.
[0003] Current research shows that for collaborative transport systems, it is common to simplify the system as a linear form for control, but this simplification will lose some control accuracy in actual control. In contrast, nonlinear model predictive control can well handle the constraints of input and output while retaining the relatively accurate nonlinear model characteristics of the system. However, due to the large amount of calculation required by nonlinear model predictive control, it is difficult to perform real-time transport control on the device. SUMMARY
[0004] The present application provides a kind of quadrotor unmanned aerial vehicle collaborative transport control method based on deep neural network, utilizes deep neural network to complete collaborative transport control task, realizes the deep neural network controller in having the control performance similar to nonlinear model predictive controller While greatly reducing the amount of calculation required, effectively solve the problem that the demand of nonlinear model predictive controller to computing power is too large and cannot be deployed on actual device.
[0005] To achieve the above object, the present application adopts the following technical solutions:
[0006] A quadrotor unmanned aerial vehicle cooperative transportation control method based on a deep neural network, comprising the following steps:
[0007] S1: establishing a dynamic equation of a quadrotor unmanned aerial vehicle cooperative transportation system;
[0008] S2: designing a nonlinear model predictive controller for the dynamic equation, and collecting flight data in a simulation environment;
[0009] S3: designing a deep neural network based on the nonlinear model predictive controller, and training the deep neural network using the collected flight data.
[0010] In the above steps, the dynamic equation of the quadrotor unmanned aerial vehicle cooperative transportation system is: , where is the inertia matrix of the cooperative transportation system, is the Coriolis matrix of the system, is the gravity vector of the system, is the generalized force vector of the system.
[0011] The nonlinear model predictive controller is: , where is the optimal control output sequence that minimizes the cost function in the prediction horizon, is the sampling time interval, is the lower limit vector of the control output, is the upper limit vector of the control output, is the running cost term, is the terminal cost term, is the prediction interval, is the number of steps of prediction, and are diagonal weight matrices, and are error vectors of generalized coordinates.
[0012] The designed neural network adopts a full connection mode between layers, and the neural network consists of an input layer, n hidden layers and an output layer.
[0013] Beneficial effects: the application provides a quadrotor unmanned aerial vehicle cooperative transportation control method based on a deep neural network, constructs a dynamic equation capable of describing the cooperative transportation of a rod-shaped load by two quadrotor unmanned aerial vehicles, designs a nonlinear model predictive controller for the dynamic equation, and collects flight data through numerical simulation, then designs a deep neural network based on the designed nonlinear model predictive controller, and trains the neural network using the collected flight data, so that the neural network can fit the control performance of the nonlinear model predictive controller. The numerical simulation results show that the neural network controller of the application can not only effectively fit the control effect of the nonlinear model predictive controller, but also significantly reduce the demand for computing resources. BRIEF DESCRIPTION OF DRAWINGS
[0014] Figure 1 is a schematic diagram of a quadrotor unmanned aerial vehicle cooperative transportation system in an embodiment of the application;
[0015] Figure 2 is a structural diagram of a deep neural network in an embodiment of the application;
[0016] Figure 3 is a curve diagram of the change of the position coordinates of the unmanned aerial vehicle in an embodiment of the application;
[0017] Figure 4 is a curve diagram of the change of the pitch angle coordinates in an embodiment of the application;
[0018] Figure 5 is a comparison diagram of the DNN and NMPC calculation time in an embodiment of the application;
[0019] Figure 6 is a curve diagram of the loss change of the neural network in the training process in an embodiment of the application. DETAILED DESCRIPTION
[0020] The application will be described in detail below in combination with the drawings and specific embodiments:
[0021] A quadrotor unmanned aerial vehicle cooperative transportation control method based on a deep neural network, specifically comprising the following steps:
[0022] Assuming that the cable is always taut during transportation and its mass can be ignored, the dynamic equation of the quadrotor unmanned aerial vehicle cooperative transportation system is established, as shown in Figure 1 The system of the embodiment is composed of two identical quadrotor unmanned aerial vehicles, a rod-shaped load and two cables, , , and are the ground coordinate system, the body coordinate system of the first quadrotor unmanned aerial vehicle, the body coordinate system of the second quadrotor unmanned aerial vehicle and the body coordinate system of the load, respectively, wherein the masses of the two unmanned aerial vehicles are equal to M, and the length of the cable is Lc , the mass and length of the rod-shaped load are m and L p , the moment of inertia of the load are I px and I pz , and are two end points of the rod-shaped load.
[0023] The position coordinates x1, y1, z1 of the first unmanned aerial vehicle, the swing angle coordinates θ 1x , θ 1y of the first cable, the swing angle coordinates α, β of the load and the swing angle coordinates θ 2x , θ 2y of the second cable are selected as the generalized coordinates, and the dynamics equation of the following system can be obtained by using the second Lagrange equation: , (1) where is the inertia matrix of the cooperative transportation system, is the Coriolis matrix of the system, is the gravity vector of the system, is the generalized force vector of the system.
[0024] The specific form of the inertia matrix is: , (2) In the above formula: ,
[0025] The specific form of the Coriolis matrix is: , (3) In the above formula: ,
[0026] The specific expression of , (4)
[0027] The specific expression of , (5) where, in the above several expressions , , , , , , , , , , and are the abbreviations of , , , , , , , , , , and , and are the short forms of and .
[0028] Based on equation (1), the discrete-time state equation of the system can be expressed as: , (6) where k is the current sampling time, is the sampling time interval, and the cost function of the nonlinear model predictive controller can be defined as: , (7) where is the running cost term, is the terminal cost term, is the prediction horizon, is the number of steps of prediction, and are diagonal weight matrices, and are the error vectors of the generalized coordinates.
[0029] At the same time, the constraints on the control output should be satisfied: , (8) where is the lower limit vector of the control output, is the upper limit vector of the control output.
[0030] Therefore, considering the generalized coordinates of the collaborative transportation system as the optimization objective, the optimization problem of the classical nonlinear model predictive control of equations (6)-(8) is as follows: , where is the optimal control output sequence that minimizes the cost function in the prediction horizon.
[0031] In order to fit the above nonlinear model predictive controller, we design Figure 2 The fully connected deep neural network shown in the figure uses a fully connected method between the neural network layers. The neural network consists of an input layer, n hidden layers and an output layer. Each hidden layer contains neurons , and Represents two quadrotors in Position information tracking error and velocity state vector at the sampling moment, express Control input at any moment.
[0032] To train a deep neural network using a neural network framework, you first load and batch the training data, define the inherited model class, and implement forward propagation. You then select a loss function and optimizer, iteratively performing forward loss calculations, backpropagation, and parameter updates in a training loop. Gradients must be manually cleared after each round. You then switch between validation and testing modes and calculate metrics such as accuracy. Finally, you save the model parameters to support subsequent deployment or transfer learning.
[0033] The proposed neural network control method is verified through numerical simulation results.
[0034] The system parameters are defined as follows: the mass of the quadrotor drone and the payload are and , the lengths of the two cables and the load are and The moment of inertia of the load is and .
[0035] The main parameters of the nonlinear model predictive controller are set as follows: the sampling time interval is , the prediction interval is The total numerical simulation time is , the simulation interval is , the running loss diagonal weight matrix is , the terminal loss diagonal weight matrix is , the upper limit of the control input is , the lower limit of the control input is .
[0036] The main parameters of the deep neural network are set as follows: the number of hidden layers of the neural network is , the number of neurons in each hidden layer is , the number of neurons in the input layer is 12, the number of neurons in the output layer is 6, and the activation function is , the loss function is , the learning rate is , the batch training sample size is 32, the training round is 2000, and the L2 regularization strength is , and the optimization algorithm is Adam.
[0037] The control effects of the nonlinear model predictive controller and the neural network controller are compared as shown in Figure 3 and Figure 4 , wherein the nonlinear model predictive control is represented by the abbreviation NMPC, and the neural network controller is represented by the abbreviation DNN. Figure 3 and Figure 4 The blue dotted line and the red solid line in
[0038] Figure 3 The curves of the position coordinates of the two unmanned aerial vehicles of the cooperative transportation system controlled by the NMPC and the DNN are given. The average rise time of the NMPC controller is about 1.30 s, the average peak time is about 1.94 s, the average overshoot is less than 8%, the average adjustment time is about 2.50 s, and the steady-state error is kept within 1%. The average rise time of the DNN controller is about 1.28 s, the average peak time is about 1.92 s, the average overshoot is less than 10%, and the average adjustment time is about 2.50 s. Figure 4 The curves of the pitch angle coordinates of the cooperative transportation system controlled by the NMPC-PD and the DNN-PD are given, and it can be seen that both controllers can effectively suppress the load swing. Figure 5 The comparison chart of the calculation time of the two controllers is given, Figure 6 is the loss change curve in the neural network training process, and the numerical simulation results show that the neural network controller can effectively fit the control effect of the nonlinear model predictive controller while significantly reducing the demand for computing resources.
[0039] The above only describes the preferred embodiments of the present application, and it should be pointed out that those skilled in the art can make corresponding changes and adjustments to the technology of the present application without departing from the basic principles of the present application, and these changes and adjustments all belong to the protection scope of the present application.
Claims
1. A method for cooperative transportation control of a quadrotor drone based on deep neural network, characterized in that: The following steps are involved: S1: Establish the dynamic equations of the quadrotor UAV cooperative transportation system; S2: Design a nonlinear model predictive controller for the dynamic equations and collect flight data in a simulation environment; S3: Designing a deep neural network based on the nonlinear model predictive controller, and training the deep neural network using the collected flight data.
2. The method for controlling the coordinated transport of a quadrotor drone based on a deep neural network according to claim 1, It is characterized in that The dynamic equation of the quadrotor UAV cooperative transport system is obtained using the second-kind Lagrange equation: Among them, M c (q) is the inertia matrix of the cooperative transport system, is the Coriolis force matrix of the system, G(q) is the gravity vector of the system, and U is the generalized force vector of the system.
3. The method for controlling the coordinated transport of a quadrotor drone based on a deep neural network according to claim 2, It is characterized in that The inertia matrix M of the cooperative transport system c The specific form of (q) is: Where: m c11 =m c22 =m c33 =2M+m, Among them, M is the mass of the UAV, L c The length of the cable is m and L p are the mass and length of the rod load, I px and I pz are the moment of inertia of the load respectively.
4. The method for controlling the coordinated transport of a quadrotor drone based on a deep neural network according to claim 2, wherein: The Coriolis force matrix of the system The specific form is: Where:
5. The method for controlling the coordinated transport of a quadrotor drone based on a deep neural network according to claim 2, wherein: The gravity vector of the system G(q) The specific expression is:
6. The method for controlling the coordinated transportation of a quadrotor drone based on a deep neural network according to claim 2, wherein: The specific expression of the generalized force vector U of the system is: Among them, S x1 ,C x1 ,S y1 ,C y1 ,S α ,C α ,S β ,C β ,S x2 ,C x2 ,S y2 and C y2 sinθ respectively 1x ,cosθ 1x ,sinθ 1y ,cosθ 1y ,sinα,cosα,sinβ,cosβ,sinθ 2x ,cosθ 2x ,sinθ 2y and cosθ 2y Abbreviation, l p and l c L p and L c The abbreviation of θ 1x ,θ 1y is the swing angle coordinate of the first cable, α and β are the swing angle coordinates of the load, θ 2x ,θ 2y is the swing angle coordinate of the second cable.
7. The method for controlling the coordinated transportation of a quadrotor drone based on a deep neural network according to claim 1, wherein: The nonlinear model predictive controller is: q(k+δ t )=f(q(k),u * (k)), in min in * (k)≤u max . Among them, u * (k) is the optimal control output sequence that minimizes the cost function in the prediction time domain, δ t is the sampling time interval, u min is the lower limit vector of the control output, u max is the upper limit vector of the control output, e q (i|k) T Pe q (i|k) is the running cost item, e q (N p |k) T Fe q (N p |k) is the terminal cost term, N p is the prediction interval, i is the number of prediction steps, P>0 and F>0 are diagonal weight matrices, e q (i|k) and e q (N p |k) is the error vector of generalized coordinates.
8. The method for controlling the coordinated transportation of a quadrotor drone based on a deep neural network according to claim 1 or 7, wherein: The designed neural network consists of an input layer, n hidden layers and an output layer, and the layers are fully connected.
9. The method for controlling the coordinated transportation of a quadrotor drone based on a deep neural network according to claim 8, characterized in that: Use neural network frameworks to train deep neural networks: First, load and batch the training data, define the inherited model class, and implement forward propagation. Select the loss function and optimizer, and iteratively perform forward loss calculation, backpropagation, and parameter updates in the training loop. Manually clear the gradients after each round. Switch to validation / test mode and calculate the accuracy. Finally, save the model parameters to support subsequent deployment or transfer learning.