Comparison of feature extraction from wavelet packet based on reconstructed signals versus wavelet packet coefficients for fault diagnosis of rotating machinery

This is the authors' abstract. We don't add key points for this paper.

Vibration signals from rotating machines are usually nonlinear and nonstationary. Time frequency techniques are suitable for analyzing this type of signals. Wavelet analysis is one of the most powerful methods in this regards. Therefore, wavelet analysis is used extensively for diagnosis of nonlinear and nonstationary signals. Faults in rotating machines show their effects in certain frequency bands. In this research the features extracted from reconstructed signals from wavelet packets were compared to features extracted from wavelet packet coefficients. It is shown that reconstructed signals act better for fault diagnosis than wavelet packet coefficients. To support our claim one example is designed that justifies our hypothesis. To evaluate our hypothesis in real world practical situations, health condition monitoring of a motorcycle gearbox has been considered. In this practical situation wavelet coefficients and reconstructed signals from wavelet packet coefficients extracted from signals acquired from gearbox housing were compared. Mahalanobis distance (MD) is employed to evaluate the significance of the extracted features. It is shown that features extracted from reconstructed signals are more suitable than features extracted from wavelet packet coefficients.