Abstract

In complex systems such as aircraft engines, system reliability and adequate monitoring is of high priority. The performance of all physical systems degrades over time due to aging, the working and environmental conditions. Considering both time and safety, it is important to predict the system health condition in future in order to be able to assign a suitable maintenance policy. Towards this end, two artificial intelligence based methodologies are proposed and investigated in this thesis. The main objective is to predict degradation trends by studying their effects on the engine measurable parameters such as the temperature and pressure at critical points of a gas turbine engine.

The first proposed prognostic scheme for the gas turbine engine is based on a recurrent neural networks (RNN) architecture. This closed-loop architecture enables the network to learn the increasing degradation dynamics using
the collected data set. Training the neural networks and determining the suitable number of network parameters are challenging tasks. The other challenge associated with the prognostic problem is the uncertainty management. This is inherent in such schemes due to measurement noise and the fact that one is trying to project forward in time. To overcome this problem, upper and lower prediction bounds are defined and obtained in this thesis. The two bounds constitute a prediction band which helps one not to merely depend on the single point prediction. The prediction bands along with the prediction error statistical measures, allow one to decide on the goodness of the prediction results.

The second prognostic scheme is based on a nonlinear autoregressive with exogenous input (NARX) neural networks architecture. This recurrent dynamical structure takes advantage of both features which makes it easy to manage the main objective. The network is trained with fewer examples and the prediction errors are lower as compared to the
first architecture. The statistical error measures and the prediction bands are obtained for this architecture as well.

In order to evaluate and compare the prediction results from the two proposed neural networks a metric known as the normalized Akaike information criterion (NAIC) is applied in this thesis. This metric takes into account the prediction error, the number of parameters used in the neural networks architecture and the number of samples in the test data set. A smaller NAIC value shows a better, more accurate and more effective prediction result. The NAIC values are found for each case and the networks are compared at the end of the thesis.

Neural networks performance is based on the suitability of the data they are provided with. Two main causes of engine degradation are modelled in this thesis and a SIMULINK model is developed. Various scenarios and case studies are presented to illustrate and demonstrate the effectiveness of our proposed neural networks based prognostic approaches. The prognostic results can be employed for the engine health management purposes. This is a growing and an active area of research for the aircraft engines where only a few references exist in the literature.