On-Line Reliability Estimation of Physical Systems Using Neural Networks and Wavelets

Individual component reliability can often be estimated from degradation signals. In this paper, we examine the utility of the wavelet transform in pre-processing degradation signals for on-line reliability estimation. Wavelet preprocessing facilitates examination of degradation signals in both the timeand frequency-domains, simultaneously. Neural networks are used for forecasting the degradation signals (or a transformation thereof) and estimating the likelihood that these signals would exceed a pre-determined critical plane representative of unit failure in the immediate future. This leads to an on-line estimate for individual unit reliability. The proposed method is applied for analyzing degradation signals collected from a vertical CNC drilling machine using drill-bits. The degradation signals, force and torque, were collected as the drillbits were destructively tested.