Optimization of spectral kurtosis-based filtering through flow direction algorithm for early fault detection

Abstract

This research focuses on developing a denoising filter that effectively enhances subtle non-stationarities within signals. Initially, the spectral kurtosis has been calculated from each frequency bin of the spectrogram (setting initial values) which is further optimized by the flow direction algorithm (FDA) to obtain a vector of optimized spectral kurtosis as the classical spectral kurtosis as a data-driven estimator may be not optimal in the sense of informative signal extraction for some specific signals with complex time–frequency structure. Due to this, the vectors obtained from optimized spectral kurtosis for a given frequency bin could be even zero if they do not support the impulsiveness of the output signal. In this way, it allows not to use any thresholding that is always complicated. The vector of the optimized spectral kurtosis thus obtained has been used to design the filter that can extract the impulsive components from the signal.