Abstract:
Flight stability derivatives are critical parameters in the design and control of aircraft, influencing overall stability and performance. Accurate estimation of these derivatives is essential
for developing advanced flight control systems, aircraft design, and pilot training. Traditional
methods for estimating flight stability derivatives often rely on complex mathematical models
and wind tunnel testing, which can be costly and time-consuming.
This thesis investigates the application of neural networks to efficiently and accurately estimate
flight stability derivatives. Neural networks, with their ability to model nonlinear relationships
and handle large datasets, offer a promising alternative to traditional methods. Using experimental and simulation data, this research aims to develop a robust framework for applying System
Identification techniques, which refer to a set of mathematical tools and algorithms used to build
mathematical models of dynamic systems, to fifth-generation fighter aircraft.
Different neural network architectures, including FFNN, RNN, LSTM, and GRU, were tested
and analyzed on static and dynamic NASA datasets. These approaches showed promising results, with LSTM performing the best based on established evaluation metrics such as Mean
Squared Error (MSE) and Root Mean Squared Error (RMSE). Neural networks, in general, also
performed better than traditional methods currently used in the industry. A comparative study
with Kalman Filter showed that the neural networks were not only more efficient in terms of
accuracy but also computationally faster by at least 60 times.
