This dissertation investigates the effectiveness of a two stage fault identification methodology for rotating machines operating under non-stationary conditions with the use of a single vibration transducer. The proposed methodology transforms the machine vibration signal into a discrepancy signal by means of smart feature selection and statistical models. The discrepancy signal indicates the angular position and relative magnitude of irregular signal patterns which are assumed to be indicative of gear faults. The discrepancy signal is also independent of healthy vibration components, such as the meshing frequency, and effects of fluctuating operating conditions. The use of the discrepancy signal significantly reduces the complexity of fault detection and diagnosis. The first stage of the methodology involves extracting smart instantaneous operating condition specific features, while the second stage requires extracting smart instantaneous fault sensitive features. The instantaneous operating condition features are extracted from the coefficients of the low frequency region of the STFT of the vibration signal, since they are sensitive to operating condition changes and robust to the presence of faults. Then the sequence of operating conditions are classified using a hidden Markov model (HMM). The instantaneous fault features are then extracted from the coefficients in the wavelet packet transform (WPT) around the natural frequencies of the gearbox. These features are the converse to the operating condition features,since they are sensitive to the presence of faults and robust to the fluctuating operating conditions. The instantaneous fault features are sent to a set of Gaussian mixture models (GMMs), one GMM for each identified operating condition which enables the instantaneous fault features to be evaluated with respect to their operating condition. The GMMs generate a discrepancy signal, in the angular domain, from which gear faults may be detected and diagnosed by means of simple analysis techniques. The proposed methodology is validated using experimental data from an accelerated life test of a gearbox operated under fluctuating load and speed conditions.
Dissertation (MEng)--University of Pretoria, 2015.