Fault feature extraction using nonlinear wavelet transform

Because one wavelet basis is adopted at every level in classical wavelet transform that can not ideally match the local characteristics of signals, some useful information of original signals is lost in denoised signals. In order to overcome the mentioned limitation, a pre-processing method based on nonlinear wavelet transform for vibration signals is adopted by using second generation wavelet transform (SGWT). By virtue of the property of predictor and updater independent each other in SGWT, an optimal predictor is selected for a transforming sample according to the selection criterion of minimizing the squared error for prediction. Consequently, the selected predictor can always fit the local characteristics of the signals. The simulations showed that the proposed method could overcome the disadvantage of classical denoising approach that lose local information of original signals. It not only can filter noise from original signals effectively, but also can hold local characteristics of original signals in the denoised signals. Fault features were successfully extracted from vibration signal for further diagnosis in a power plant by taking the proposed method as a preprocessing tool.