Effective fault diagnosis techniques are important to ensure that expensive assets such as
wind turbines can operate reliably. Vibration condition monitoring data are rich with information
pertaining to the dynamics of the rotating machines and are therefore popular for
rotating machine diagnostics. However, vibration data do not only contain diagnostic information,
but operating condition information as well. The performance of many conventional
fault diagnosis techniques is impeded by inherent varying operating conditions encountered in
machines such as wind turbines and draglines. Hence, it is not only important to utilise fault
diagnosis techniques that are sensitive to faults, but the techniques should also be robust to
changes in operating conditions.
Much research has been conducted to address the many facets of gearbox fault diagnosis
e.g. understanding the interactions of the components, the characteristics of the vibration
signals and the development of good vibration analysis techniques. The aforementioned
knowledge, as well as the availability of historical data, are regarded as prior knowledge (i.e.
information that is available before inferring the condition of the machine) in this thesis.
The available prior knowledge can be utilised to ensure that e ective gearbox fault diagnosis
techniques are designed. Therefore, methodologies are proposed in this work which
can utilise the available prior knowledge to e ectively perform fault diagnosis, i.e. detection,
localisation and trending, under varying operating conditions. It is necessary to design di erent methodologies to accommodate the di erent kinds of historical data (e.g. healthy
historical data or historical fault data) that can be encountered and the di erent signal analysis
techniques that can be used.
More speci cally, a methodology is developed to automatically detect localised gear damage
under varying operating conditions without any historical data being available. The
success of the methodology is attributed to the fact that the interaction between gear teeth
in a similar condition results in data being generated which are statistically similar and this
prior knowledge may be utilised. Therefore, a dissimilarity measure between the probability
density functions of two teeth can be used to detect a gear tooth with localised gear damage.
Three methodologies are also developed to utilise the available historical data from a
healthy machine for gearbox fault diagnosis. Firstly, discrepancy analysis, a powerful novelty
detection technique which has been used for gear diagnostics under varying operating conditions,
is extended for bearing diagnostics under varying operating conditions. The suitability
of time-frequency analysis techniques and di erent models are compared for discrepancy analysis
as well. Secondly, a methodology is developed where the spectral coherence, a powerful
second-order cyclostationary technique, is supplemented with healthy historical data for fault
detection, localisation and trending. Lastly, a methodology is proposed which utilises narrowband
feature extraction methods such as the kurtogram to extract a signal rich with novel
information from a vibration signal. This is performed by attenuating the historical information
in the signal. Sophisticated signal analysis techniques such as the squared envelope
spectrum and the spectral coherence are also used on the novel signal to highlight the bene ts
of utilising the novel signal as opposed to raw vibration signal for fault diagnosis.
Even though a healthy state is the desired operating condition of rotating machines, fault
data will become available during the operational life of the machine. Therefore, a methodology,
centred around discrepancy analysis, is developed to utilise the available historical fault
data and to accommodate fault data becoming available during the operation of the machine.
In this investigation, it is recognised that the machine condition monitoring problem
is in fact an open set recognition problem with continuous transitions between the healthy
machine condition and the failure conditions. This is explicitly incorporated into the methodology
and used to infer the condition of the gearbox in an open set recognition framework.
This methodology uses a di erent approach to the conventional supervised machine learning
techniques found in the literature.
The methodologies are investigated on numerical and experimental datasets generated under varying operating conditions. The results indicate the bene ts of incorporating prior
knowledge into the fault diagnosis process: the fault diagnosis techniques can be more robust
to varying operating conditions, more sensitive to damage and easier to interpret by a
non-expert. In summary, fault diagnosis techniques are more e ective when prior knowledge