Modélisation des performances du moteur General Electric CF34-8C5B1 de l’avion CRJ700 à partir des réseaux de neurones et des tests en vol

Abstract

A new methodology was developed and then validated in order to identify a static performance model of the General Electric CF34-8C5B1 Turbofan using two successive Feedforward Neural Networks and flight tests. The flight tests were carried on a Level D flight simulator of the CRJ700 for the aircraft flight envelope including major flight phases such as takeoff, climb, cruise and descent. The identification process was divided in two steps. In the first step, the following engine design parameters were estimated : Engine Pressure Ratio EPR, Fan Pressure Ratio FRR, Overall Pressure Ratio OPR, Inlet Turbine Temperature ITT, Fan speed N1 and Core speed N2 using a first neural networks ECS-NET. This first step aimed to reproduce the Engine Control System by modelling the thrust ratings and each engine component characteristics. Then, the second step consisted in predicting the engine performance parameters, such as the fuel flow Wf and the thrust Fn using a second neural network ENG-NET. The model inputs were the altitude, the Mach number and the atmosphere conditions such as ambient pressure, ambient temperature and air density. The identification process was successfully carried on Matlab/Simulink engineering software and gave excellent performances. The neural networks were trained using the Bayesian regularization algorithm, which was able to identify a non-linear system from a minimum amount of data. Several output combinations for the ECS-NET model were proposed, which demonstrated the ability of the neural networks to adapt according to the given set of parameters. The validation phase was performed using a comparative analysis between the model, and the reference aircraft outputs. To do this phase, a statistical analysis was processed for the overall flight tests, and then several simulation scenarios were studied, including an engine degradation scenario during the cruise phase. The results have shown that the models have an excellent ability of prediction, with a relative error less than 2% and a standard deviation less than 5% for each output, as required by the FAA.

Date	13 Jun 2021
Original language	French
Awarding Institution	École de technologie supérieure
Supervisor	Ruxandra Botez (Supervisor)