Perpustakaan Digital ITB

The primary function of a flight control system is to maintain the stability and manage the attitude and manoeuvre of the aircraft. In developing a flight control system, issues such as the dynamic behaviour of the aircraft and operational aspects, must be considered, especially since in many cases it will work together with human a controller. To tackle these issues, a deep learning neural network scheme is chosen due to its ability to consider not only analytical or numerical model based information, but also empirical information from good practical implementation in controlling an aircraft, for constructing the controller. This information is exploited as the datasets for learning process in order to set the parameters of the controller. In this work, the numerical Research Civil Aircraft Model (RCAM), equipped with a linear multivariable controller, is simulated to generate dataset for training a configuration of Neural Network (NN). The generated data focuses on a condition in cruise flight, where the controller works to maintain the stability while also managing the variation of some flight variables of the aircraft. With this data a Nonlinear Autoregressive with Exogenous Inputs (NARX) neural network is trained and implemented into the system. A NARX net is chosen for its capability in handling nonlinear data with the use of both a sigmoid and linear activation function in its architecture. The trained NN model, which replicates the simulated controller, is then applied to the RCAM model. Results concluded that the neural network is able to perform well both inside and outside its training scope for some intended manoeuvres, within a predefined limit. Future endeavours should consider a larger dataset with a variety of conditions and the use of actual flight recorder data to better reflect a real flight scenario.