Pinguet, Jérémy (Safrran Electronics & Defense), Feyel, Philippe (Safran Electronics and Defense), Sandou, Guillaume (École supérieure d'électricité)

A Neural Autopilot Training Platform Based on a Matlab and X-Plane Co-Simulation

Scheduled for presentation during the Regular Session "Neural Networks" (FrB3), Friday, June 18, 2021, 11:40−12:00, Edessa

2021 International Conference on Unmanned Aircraft Systems (ICUAS), June 15-18, 2021, Athens, Greece

This information is tentative and subject to change. Compiled on April 19, 2024

Abstract

The main objective of this paper is to describe a tool for the aircraft autopilot deployment only based on a flight database. Flight simulators such as X-Plane turn out to be powerful and efficient tools for creating such database and controller experimentation. The paper outlines the development of a co-simulation framework between Matlab and X-Plane using the User Datagram Protocol (UDP). The flight data collected during a first step are then used for the training of neural controllers. The approach is based on the neural network imitation ability to learn the piloting skills implicitly stored in the dataset. Also, in order to include fault-tolerant control, a Neural Multiple Model Adaptive Control (NMMAC) based on previously learned networks is implemented. This architecture consists of a bank of local controllers and a switching logic using a bank of estimators. As an illustration of the proposed platform, it is assumed that the airspeed is unmeasured for the flight director. A neural guidance autopilot based NMMAC is therefore performed on different airspeed values. Experiments show that the designed neural autopilot can successfully track both heading and altitude reference signals, while the method is not restricted to this scope.