No description available.

Condition monitoring of rotating shafts is gaining importance in industry due to the need to increase machine reliability and decrease the possible loss of production due to machine breakdown. In this work, the use of vibration signals for the detection of a crack within a shaft was investigated. The research involved the measurement of vibration signals during laboratory tests on a long rotating shaft rig. The focus of the experimental work was on the effect of cracks on the dynamics and the initiation and growth of cracks in the shaft. Measurements were taken from the shaft system both with simulated cracks (notches) cut at 45° and 90° to the shaft axis and with real propagating cracks initiated by a pre-crack cut. All defects were located at the mid- point along the shaft. The vibration responses and stresses were measured for different depths of crack. The vibration responses of the three different defects were compared using PSDs of the data to identify the change in position and magnitude of the peaks in the spectrum under each defect. Experiments to study the effect of defect depth at different shaft rotation speeds were also carried out. Finally, a shaft with a breathing crack (continuously opening and closing as the shaft rotates) was also studied experimentally, with the crack growing under normal steady state operating conditions. After completing the experiment work, the shaft was broken and the type of fracture studied. The results for both simulated and actual crack growth showed that vibration frequencies decreased as a crack progressed, indicating the possibility of using the vibration signal for crack detection. A significant relationship was found between the stage of crack growth and the vibration results. A finite element (FE) model was constructed to explore the relationship between the natural frequencies and crack depth and position along the shaft and to explain and validate the results of the experimental work. The FE model showed similar trends to the experimental results and also allowed the effect of different crack positions to be explored. The PSD data was fed into an artificial neural network after a feature extraction procedure was applied to significantly reduce the quantity of data whilst at the same time retaining the salient information. Such an approach results in a considerably reduced training time for the network due to the reduced complexity. The proposed scheme was shown to successfully identify the different defect levels. This method greatly enhances the capacity of an automated diagnostic process by linking increased capability in signal analysis to the predictive capability of the artificial neural network.