Multiresolution error detection on early fatigue cracks in gears

Abstract

In this paper, we focused on early fatigue cracks in gears in a helicopter transmission test rig. The time synchronous average signal is transformed to different scale using wavelet transform. The wavelet vanishing moments can characterize the local signal singularities defined as local Holder exponent. At each level, one can extract from the signal its regularity: approximate the wavelet-transformed signal using the linear models: Autoregressive Moving Average (ARMA). The residual error is computed at each level. The final phase is the feature extraction from the residual error at each scale. The probability density function of the residual error is expanded into Hermite polynomial. The coefficients of this expansion are used as a feature vector for detection/estimation of the early fatigue cracks in gears. One can track the nonstationarity signal embedded in the residual error at each scale.