Fault diagnostic methods aim to recognize when a fault exists on a system and to identify
the failures which have caused it. The fault symptoms are obtained from readings of sensors
located on the system. When the observed readings do not match those expected then a fault
can exist. Using the detailed information provided by the sensors a list of the failures that are potential causes of the symptoms can be deduced. In the last decades, fault diagnostics
has received growing attention due to the complexity of modern systems and the consequent
need of more sophisticated techniques to identify failures when they occur. Detecting the
causes of a fault quickly and efficiently means reducing the costs associated with the system
unavailability and, in certain cases, avoiding the risks of unsafe operating conditions.
Bayesian Belief Networks (BBNs) are probabilistic graphical models that were developed for
artificial intelligence applications but are now applied in many fields. They are ideal for
modelling the causal relations between faults and symptoms used in fault diagnostic processes.
The probabilities of events within the BBN can be updated following observations
(evidence) about the system state.
In this thesis it is investigated how BBNs can be applied to the diagnosis of faults on a
system with a model-based approach. Initially Fault Trees (FTs) are constructed to indicate
how the component failures can combine to cause unexpected deviations in the variables
monitored by the sensors. The FTs are then converted into BBNs and these are combined
in one network that represents the system. The posterior probabilities of the component
failures give a measure of which components have caused the symptoms observed. The technique
is able to handle dynamics in the system introducing dynamic patterns for the sensor
readings in the logic structure of the BBNs.
The method is applied to two systems: a simple water tank system and a more complex fuel
rig system. The results from the two applications are validated using two simulation codes
in C++ by which the system faulty states are obtained together with the failures that cause
them. The accuracy of the BBN results is evaluated by comparing the actual causes found
with the simulation with the potential causes obtained with the diagnostic method.
A Doctoral Thesis. Submitted in partial fulfillment of the requirements for the award of Doctor of Philosophy of Loughborough University.