Abstract

In this thesis a new approach for jet engine Fault Detection and Isolation (FDI) is proposed using ensemble neural networks. Ensemble methods combine various model predictions to reduce the modeling error and increase the prediction accuracy. By combining individual models, more robust and accurate representations are almost always achievable without the need of ad-hoc fine tunings that are required for single model-based solutions.

For the purpose of jet engine health monitoring, the model of the jet engine dynamics is represented using three different stand-alone or individual neural network learning algorithms. Specifically, a dynamic multi-layer perceptron (MLP), a dynamic radial-basis function (RBF) neural network, and a dynamic support vector machine (SVM) are trained to individually model the jet engine dynamics. The accuracy of each stand-alone model in identification of the jet engine dynamics is evaluated. Next, three ensemble-based techniques are employed to represent jet engine dynamics. Namely, two heterogenous ensemble models (an ensemble model is heterogeneous when different learning algorithms (neural networks) are used for training its members) and a homogeneous ensemble model (all the models are generated using the same learning algorithm (neural network)). It is concluded that the ensemble models improve the modeling accuracy when compared to stand-alone solutions. The best selected stand-alone model (i.e the dynamic radial-basis function neural network in this application) and the best selected ensemble model (i.e. a heterogenous ensemble) in term of the jet engine modeling accuracy are selected for performing the FDI study.

Engine residual signals are generated using both single model-based and ensemble-based solutions under various engine health conditions. The obtained residuals are evaluated in order to detect engine faults. Our simulation results demonstrate that the fault detection task using residuals that are obtained from the ensemble model results in more accurate performance. The fault isolation task is performed by evaluating variations in residual signals (before and after a fault detection flag) using a neural network classifier. As in the fault detection results, it is observed that the ensemble-based fault isolation task results in a more promising performance.