Translation Invariant Features from Cascaded Wavelet and Fourier Transforms for Bearing Fault Pattern Recognition Using Artificial Neural Network

Rotary machine fault classification from vibrations requires robust feature extraction and enhancement procedures for transient and steady-state fault signatures. Accurate fault pattern classification relies on the quality of features extracted from the fault patterns. Fourier transform (FT) and wavelet transform (WT) based methods have largely been used for extraction of these features. FT performs well with non-stationary vibrations to provide translation invariant spectral features which can readily be used as input for classifier but belittles the spectral amplitudes of time-domain transients because of unmatched window size. WT, in contrast, deals well with transient’s amplitude calculations from non-stationary vibrations because of signal decomposition into several frequency sub-bands but lacks in readily providing translation invariant features. As WT can better augment features and FT can readily provide translation-invariant spectral features suited for artificial neural network (ANN) classifier, therefore, this paper proposes a cascaded WT and FT based features extraction method for improved fault pattern recognition. The efficacy of proposed work is evaluated by comparing with existing methods. The results, under very poor SNR of -10dB, show that cascaded WT and FT based augmented and translation invariant features with ANN surpasses existing methods in classification accuracy under given conditions.