Neural Networks and Support Vector Regression for the CRJ-700 Longitudinal Dynamics Modeling

Abstract

This paper presents a new methodology to identify the aerodynamic coefficients and predict the short-period and phugoid dynamics of an aircraft using two types of supervised learning models: artificial neural network (ANN), and support vector regression (SVR). The study was validated on the CRJ 700 regional jet. Simulated flight tests data were collected during various maneuvers performed on a Level D CRJ-700 Virtual Research Simulator (VRESIM) designed by CAE and Bombardier. Level D is the highest qualification for flight dynamics and propulsion models given by the FAA. Both ANN and SVR models were trained using the data collected from the VRESIM to develop multidimensional models capable of predicting the aerodynamic coefficients of the aircraft for any conditions in the flight envelope, defined by altitude, speed, weight, and center-of-gravity position. The choice of solvers and the optimization of hyperparameters are detailed for both types of models. These models were validated by comparing predicted flight parameters with experimental data obtained from the CRJ 700 Level D VRESIM considering the same pilot inputs. The results showed that both types of models (ANN and SVR) were able to reproduce with excellent accuracy the nonlinear behavior of the short-period and phugoid dynamics.