Condition-Based Machinery Maintenance

Three new techniques have been developed over the past 36 months that offer significant potential for on-board monitoring of mechanical systems to detect and classify incipient failures. The first is the wavelet transform, and its extension to wave packets, as a means for isolating changes in signal structure in terms of scale as well as time. The second is an extension of classical time-series analysis to multiresolution representations of a time signal, with algorithms to estimate characteristic functions of the spectrum of the signal at each resolution. The third is the maturation of artificial neural network technology to the point where it may be reliably considered as an element of an on-board diagnostic system. This effort combines all three into a comprehensive approach to incipient failure detection and classification. The front-end of our proposed classifier extracts novel feature sets offered by the wavelet transform and the multiresolution tLme series spectral estimators. These define a feature space on which a classical neural net classifier may be trained and evaluated.