Implementation of a neural network for non-linearities estimation in a tail-sitter aircraft

TL;DR

This paper presents a lightweight neural network implementation in C++ for PX4 autopilot to estimate nonlinear aerodynamic forces in tail-sitter aircraft, improving control during all flight phases.

Contribution

The work introduces a neural network-based method for real-time nonlinear force estimation integrated into an open-source autopilot system.

Findings

Neural network accurately estimates aerodynamic forces during flight.

Implementation does not significantly impact autopilot performance.

Improves control during hover, cruise, and transition phases.

Abstract

The control of a tail-sitter aircraft is a challenging task, especially during transition maneuver where the lift and drag forces are highly nonlinear. In this work, we implement a Neural Network (NN) capable of estimate such nonlinearities. Once they are estimated, one can propose a control scheme where these forces can correctly feed-forwarded. Our implementation of the NN has been programmed in C++ on the PX4 Autopilot an open-source autopilot for drones. To ensure that this implementation does not considerably affect the autopilot's performance, the coded NN must be of a light computational load. With the aim to test our approach, we have carried out a series of realistic simulations in the Software in The Loop (SITL) using the PX4 Autopilot. These experiments demonstrate that the implemented NN can be used to estimate the tail-sitter aerodynamic forces, and can be used to improve the control algorithms during all the flight phases of the tail-sitter aircraft: hover, cruise flight, and transition.