Optimal frequency band selection via stationarity testing in time frequency domain

In the last decades a plethora of signal processing tools have been proposed for the analysis of vibration signals, focusing on fault detection and diagnosis of rotating machinery. Despite the existence of numerous methodologies, there is still a need to construct more specific diagnostic algorithms. The informative frequency band is a purely frequency-domain idea, so very often the approach of spectral selectors is pursued and those are digital filters prepared in a custom way. In this paper a novel signal processing approach is proposed based on the analysis of the time-frequency domain. The vibration signal is firstly transformed to the time-frequency domain. Moreover the stationarity of the time series vectors within the discrete frequency bins of the Time-Frequency (T-F) map is statistically tested by using the Augmented Dickey-Fuller (ADF) test. A frequency band filter selector is then constructed by the aggregation of the ADF statistic values of narrow frequency bins into a vector spanning over the whole Nyquist band of the given signal.