This thesis covers the scope of one out of many ways to diagnose gearboxes, demodulating the excited vibrational signals to enhance fault detection and identification. The topic is not only of academic interest since the achievements that can be made by successful machine condition monitoring in the industry. It has a potential value that is close to be absurd, for example unplanned production stops is commonly known to be one of the worst nightmares for manufacturing companies and if one can detect faults in early stages one can improve the possibilityto plan a production stop and therefore increase the profit. Four demodulation algorithms were developed and implemented in MATLAB on data characterized by close to stationarity and distinctive energy centered around the harmonics of the gearmesh frequency. The resulting algorithms for narrowband phase and amplitude demodulation was shown to outperform Hilbert transform based phase and amplitude demodulation algorithms in gearbox diagnostics. One of the goals with the thesis was therefore reached; demodulation algorithms were developed and implemented on data. A comparison of these algorithms was done and a conclusion of which demodulation technique is superior was done. Experimental work was carried out on a test-rig and both local and distributed faults were introduced to two gearboxes, one kind of fault per gearbox. However, the data acquired from the test-rig showed severe non-stationarity and smeared spectrum properties even when angular resampling was performed and therefore a major drawback of the demodulation techniques was exploited since the methods for demodulation in this thesis are not applicable for signals with smeared spectrums. The other goal was therefore not accomplished; to distinguish a local fault from a distributed fault in data acquired by experimental work by applying the selected demodulation techniques.

Performing condition monitoring related tasks on any machinery is an essential element of their rational maintenance. Endeavours to detect an incipient fault within a system serve multiple purposes from increasing the safety of people responsible for operating the machines through decreasing the running and operational costs, allowing time to plan for the inevitable repairs and making sure that the downtime of the machine is kept to an absolute minimum. All these tasks gain extra importance in a case when machines are operated in dangerous conditions putting people's lives in potential jeopardy - for instance in the field of operating a helicopter. The robust assessment of the condition of gearboxes used by helicopters has recently been given an increased attention due to a number of accidents which followed an undetected drive train component failure. The majority of the on-board mounted condition monitoring systems use vibration response signals which are specifically processed to obtain a single number which is representative of a condition of a given monitored drive train component. Those signal processing methods are called Condition Indicators (CIs). There are a number of such CIs which are already in use and they seem to adequately indicate faults in most of the cases. However in a number of instances it has been observed that the most popular parameters like Crest Factor or FM4 failed to dependably reflect the true condition of the gear causing serious accidents, some of which resulted in a number of lives being lost. For this reason the presented research is focused on investigating the limitations of the existing CIs and designing a set of improved CIs. The development process is based on overcoming the drawbacks of thetechniques used in existing CIs combined with the intelligence gathered while analysing the acceleration vibration signals which contained a gear or a bearing fault. Five new CIs are proposed and the details of their design are documented. Both the existing and the proposed CIs are applied on the available, uncorrelated datasets. The results of the comparison show that the newly developed CIs are capable of indicating a gear or a bearing fault in a more robust and dependable fashion. Each proposed CI alone may not be the most robust indicator of the actual condition of the monitored component hence the output from all proposed CIs is combined into a single indication through use of a novel data fusion model. The Combined CI created based on the data fusion model is observed to be more robust compared to each CI alone, hence it may increase the confidence level of the decision making routine and is expected to decrease the number of false alarms. The methods of the existing CIs, the proposed CIs and the data fusion techniques as well as the results of the comparison between the different approaches are present in this thesis.