The presence of gears in most mechanical and electromechanical systems makes it mandatory to incorporate ways of monitoring their state in any condition monitoring strategy. The advantages of effective gear condition monitoring, includes increased competitiveness, for industry, via the reduction of production/manufacturing costs while maintaining a high product quality level. In addition, operator safety is of interest. Research into early detection of gear failures is still an ongoing process and is approached from either one or a combination of domains—Wigner time- frequency, time-, quefrency-, and frequency-domain. Presently, there exists no one known method to effectively diagnose all the various failures, nor to provide prognosis.

This thesis investigates the use of the fourth higher moments (kurtosis) of three statistical distributions, viz. normal, Beta and Weibull, in monitoring the condition of gears. The simulated gear failures are crack, wear, together with m oderate and severe pitting. The monitoring is approached by examining, in the time-domain, the vibration signals emanating from a hydraulic gear pump in operation. The set of signals generated by each tooth mesh is treated independently of the others. This is a departure from the more common practice of combining the signals from the meshes into a data set. The advantage of this independent technique is the increased probability of detecting an incipient fault that is lim ited to a tooth.

The characteristic parameters of the statistical distributions are estim ated using both the method of moments (MM) and maximum likelihood estimation (MLE) technique. The objective is to determine if the performance of the distributions is influenced by the estimation method adopted. This could require substantial computational time, thus the feasibility of transputer based parallelisation to reduce this time is investigated.

The results indicate that the kurtosis values of the three distributions can readily detect cracks, together with moderate and severe pits. The distributions can be listed in order of performance, from best to good, as Beta, Weibull, and normal. The Beta distribution shows a better performance when the parameters are estimated using MLE raiher than MM. This is, however, untrue with the Weibull distribution. Further, parallelisation of the monitoring process on a linear network of transputers showed considerable reduction in computational times.