Feature Detection Based on Filter Banks and Higher Order Cumulants

Based on that higher-order cumulants is insensitive to add-Gaussian noise and symmetric noise, a method for detecting feature of signal embedded in additive noise is presented, which combines a filter banks and higher-order cumulants, neither the supposition of the observed signal is Gaussian or stationary, nor a prior information about the waveform and arrive time of the observed signal is necessary. The primary principle of the proposed method is to separate the spectrum of the observed signal into narrow frequency band by using a bandpass filter banks, those subfilters which have large output energies are chosen to give a time-frequency description of the observed signal approximately, each chosen subfilter is then followed by a short-time estimation of third-order cumulant so that to suppress colored noise and detect the signal feature. Simulated and experimental results show that the method is effective and practical, even if the signal feature is submerged into noise signal completely.