Identification of ground effect and intelligent control of unmanned aerial vehicles

In the levitation and landing phases of unmanned aerial vehicles, the complicated interactions between drones and environment are difficult to be modeled, and the fast-changing motion status also makes the precise control challenging. In this paper, the dynamics of drones are first modeled and simplified. Then, an adaptive controller and deep neural network (DNN) controller are used to control the actual system with ground effect to avoid rapid levitation and hard landing. The drone is controlled manually to fly up and down to gather the data containing the relationship between disturbance and motion status, and extended Kalman filter (EKF) is applied to decrease the measurement noise. Experimental results show that more training data lead to smaller control error. The neural network enhanced controller leads to smoother levitation and landing than the adaptive controller, and they are both better than the traditional PID controller.