Adaptive signal processing techniques and condition monitoring of rotating machines

The condition monitoring of complex rotating machines using vibrational signature analysis methods has been given considerable attention in recent years. The ability to diagnose a mechanical fault is enhanced if the monitoring signal can be preprocessed to reduce the effect of unwanted noise. In this work two methods have been suggested for improving the signal to noise ratio of a diagnostic signal from rotating machines. Both methods make use of an adaptive filtering process which is based on the Least Mean Square algorithm of Widrow and Hoff. Apart from the primary input which contains the corrupted signal, both these methods make use of an auxiliary or a reference input. The first method is referred to as conventional Adaptive Noise Cancelling (ANC) and is applicable in those situations where the reference input contains noise correlated with the primary input noise. The second method which is a modified ANC is proposed for those situations where the primary and reference inputs contain correlated signals and uncorrelated or weakly correlated noises. Applying both techniques to simulated as well as actual machine data, it has been shown that conventional statistical and spectral analysis techniques can be made more effective in their diagnostic roles after the application of ANC. The second part of this research is related to the estimation of the structure of the vibrational data. The Widrow-Hoff adaptive algorithm has been used to obtain the coefficients of a linear prediction filter. The structure obtained using this method has been compared with those obtained from the conventional linear prediction, Homomorphic deconvolution and the Maximum Likelihood method. It is shown that the removal of structure from a three dimensional plot may help in the trend analysis of faults in the low frequency region.