Aircraft parameter estimation using recurrent neural networks - A critical appraisal

Abstract

Artificial neural networks provide a means of nonlinear mapping of the input-output characteristics of a system. Although the static feed forward neural networks are recently being introduced as a black-box approach to model the aerodynamic coefficients, this paper investigates applicability of dynamic neural networks, often called as Recurrent Neural Networks (RNN), to aircraft parameter estimation. It is demonstrated that the RNNs are amenable to state space model representation and hence adoptable for animation of aerodynamic derivatives retaining the model13; parametrization based on the physical insight. Several13; aspects of similarity between the RNN approach and the13; conventional aircraft parameter estimation methods such13; as output error. equation error, or extended Kalman filter13; are brought out. The proposed approach is first validated13; on simulated data and then extended to estimate from13; flight data the aerodynamic derivatives pertaining to the13; lateral-directional motion of an aircraft. Based on the13; identification results. the practical utility, advantages and limitations of the RNN approach are critically appraised 13; neurons having certain progenies the artificial neural13; networks provide a means of mapping the given input-13; output data set. There are two types of networks which13; have found some widespread applications in system13; identification. They are: i) Feed Forward Neural13; Network (FFNN) and ii) Recurrent Neural Network13; (RNN). Whereas the feed forward networks, implying a13; unidirectional flow of variables. are static, the recurrent13; neural networks incorporate an output feedback, and13; hence are suitable for dynamic systems.'13; In the field of aircraft parameter estimation, the13; equation error, output error or filter error methods have13; mostly been applied during the last three decades.'* The13; hitherto or what may be termed as conventional13; approach to aircraft parameter estimation is heavily13; based on the physical insight into the system under13; investigation. In fact the fairly well understood basic13; physical principles underlying the aircraft dynamics and13; aerodynamic forces and moment acting on an aircraft13; have contributed significantly to the highly successful13; applications of the system identification methodology