Aspects of Feed Forward Neural Network Modeling and Its Application to Lateral-Directional Flight Data.

Abstract

This report investigates in some depth various aspects of feed forward neural network as a tool for aerodynamic modeling. A study has been carried out on a typical flight data to understand the parameters of interest and their influence on FFNN modeling. Based on the study carried out, optimal values and ranges of the network tuning parameters for further FFNN applications have been provided. The lateral directional motion of the two transport aircraft, C-160 and ATTAS, is modelled using FFNN and the applicability of FFNN to the flight data is established. In general, the training phase of FFNN always yielded good results but some discrepancies were observed in the prediction and validation phase. The discrepancies are in the form of offset or biases, which are attributed to the recursive method of FFNN training and to the single bias term for multiple time segments. The future scope of this work would be to improve the training algorithm to account for different flight conditions in flight data modeling using FFNN. (orig.)