微小型飞行器的神经网络动态逆控制方法

Control Method of Dynamic Inversion with Neural Network Used for MAV

微小型飞行器(MAV)精确的数学模型很难得到,限制了单纯动态逆控制方法的使用,比例-积分-微分(PID)等传统的控制方法已不能满足要求。针对这一问题,研究了应用动态逆控制方法的新途径──神经网络动态逆。选取串接积分器的多层前向神经网络训练飞行控制系统的动态逆模型,并自适应补偿逆误差。用MATLAB的NNCTRL20工具箱并结合NNSYSID20工具箱建立了仿真系统。升降舵和方向舵联合控制转弯。用PID控制器、近似神经网络动态逆模块、在线神经网络补偿器构建了飞行控制系统。仿真结果表明,神经网络动态逆有较强的鲁棒性、稳定性和指令跟随能力,比PID更适合于微小型飞行器的姿态控制。

The accurate mathematic model of micro air vehicle （MAV） can hardly be gotten. The use of control method of pure dynamic inversion is limited. The traditional control methods, such as proportion-integral-derivative （PID） etc., cannot satisfy all the requirements. Aiming at this condition, a new control method of dynamic inversion with neural network was studied and some models of neural network used in dynamic inversion were analyzed. Multilayer feed-forward neural network with integrators was chosen to train the dynamic inversion model and the according compensated error of the model. The simulator was established by combining NNCTRL20 and NNSYSID20 in MATLAB. Rudder and elevator were used to control the swerve; elevator was differential and could associate with the rudder. A new flight control system of the MAV was composed to two PID controllers, an approximate dynamic inversion block, and an adaptive on-line neural-network compensator. The system was driven by the outputs of a reference model block. Simulation results demonstrate that the control method has strong robustness, stability and capability of following commands. Compared with PID, this control method adapts more to attitudes control of the MAV.