Trajectory flight-time prediction based on machine learning for unmanned traffic management

This paper describes the study conducted to predict the trajectory flight-time of a drone adopting a machine learning approach. The proposed method has been carried out developing a feedforward neural network to estimate the flight-time needed by the drone to perform a selected corner of a designed path. To acquire a consistent database for the neural network training several reference corner paths have been flown by a test drone. The reference corners have fixed side length and different turning angle. Neural network input parameters are the corner angle, relative orientation and intensity of wind. From the telemetry analysis the flight-time to fly the corner path has been computed and employed to train the neural network. The Levenberg-Marquardt algorithm and the Bayesian Regularization backpropagation algorithm have been exploited as training functions, analyzing several neural network architectures with a different number of hidden layers and neurons. At the end, the neural networks that are characterized by the best training and test performance have been selected for each training function. Stating the trained network, a generic path has been planned to test the proposed method. The error between the estimated flight-time and the real flight-time from the drone telemetry has been evaluated.