Fast parametric time-frequency modeling of nonstationary signals

This paper studies the approximation of nonstationary signals from natural life systems in time-frequency plane using a type of four-parameter Gabor atoms. These four parameters, i.e. the dilation, chirprate, modulation and translation, have clear physical meanings and to optimize these parameters is an extremely difficult task. In this paper, a fast procedure is introduced without explicitly exploring these parameters over the continuous search space. Here, these four parameters together with a phase parameter for real signal are assigned with random values across their full ranges, creating a large library of candidate Gabor atoms. Then a fast algorithm is used to select atoms from the library that best approximate the nonstationary signal. The computational complexity of the method is only linearly related with the library size, the number of signal points and the number of used atoms. The proposed method is applied to several benchmark problems from life systems, including the bat signals, EEG signals and speech signals. The simulation results confirm its efficacy.