Neural Network-based UAV System Identification from Sparse Flight Test Data

Abstract

System identification is a powerful tool enabling to build of a "digital twin" of a system from real data for implementing controller design. Well-known models using time-based or frequency-based approaches provide reasonable analytic solutions to identify system parameters, primarily offline. The NN-based strategies have the potential to provide online solutions as they require relatively lower computational effort when training is completed in exchange for the vast amount of data. This paper here aims to provide an understanding of whether the NN-based system identification approach with the same amount of data in comparison to other well-known methodologies provides sufficient accuracy. We have chosen Ordinary Least Square (OLE) Method, which is based on regression analysis, and Output Error Method (OEM), using maximum likelihood functions to compare with NN-based identification. To gather actual flight data, we have conducted flight tests with predefined maneuvers for DEHA platform, a small-size jet-propellent UAV. Then we employed these three system identification methodologies for accuracy comparison. Comparative results showed us that NN-based system identification provides enough accurate results for a small UAV platform and is seen as a candidate for online implementations such as fault identification, online adaptive control, etc.