A Novel Neural Network Model Applied to Modeling of a Tandem-Wing Quadplane Drone

This research focuses on modeling one of the Quadplane flight phases: a hover state, similar to a regular Quadcopter hovering. The process is highly non-linear, and additionally, there are more phenomena to take into account - it is related to air turbulence around the wings and the fuselage. This work thoroughly studies the effectiveness of various types of neural networks to model the drone using the data recorded from real free-flight experiments. Finally, we introduce a novel type of neural-based model: the Feature-Sequence-To-Sequence (fseq2seq) Recurrent Neural Network Model. The new Model has interesting features: the input-data-driven initialization of RNN's internal states and a simplification of the input layer (significant reduction of used neurons' weights). We demonstrate that the new network outperforms all classic model types.