The main objective of this thesis is to design a fault detection and isolation (FDI)
scheme for the aircraft jet engine using dynamic neural networks. Toward this end
two different types of dynamic neural networks are used to learn the engine dynamics.
Specially, dynamic neural model (DNM) and time delay neural network (TDNN) are
utilized. The DNM is constructed by using dynamic neurons which utilize infinite
impulse response (IIR) filters to generate dynamical behaviour between the input
and output of the network. On the other hand, TDNN uses several delays associated
with the inputs of the neurons to achieve a dynamic input-output map. We have
investigated the fault detection performance of each structure. A bank of neural
networks consisting of a set of 12 networks that are trained separately to capture
the dynamic relations of all the 12 engine parameters are considered in this study.
The results show that certain engine parameters have better detection capabilities
as compared to the others. Moreover, the fault detection performance was improved
by introduction of the concept of "enhanced fault diagnosis scheme" which employs
several networks and monitors several engine parameters simultaneously to enhance
and improve the accuracy and performance of the diagnostic system.
The fault isolation task is accomplished by using a multilayer perception (MLP)
network as a classifier. The concept behind the isolation is motivated by the fact that
there is a specific map between the residuals of different networks and a particular
fault scenario. We show that the MLP has good capability in learning this map and
isolates the faults that are occurring in the jet engine. To demonstrate our diagnostic
scheme capabilities, 8 different fault scenarios are simulated and according to the
simulation results, our proposed FDI scheme represents a promising tool for fault
detection as well as fault isolation requirements.
iii