Adaptive short-time Fourier transform based on reinforcement learning

Abstract

Short-time Fourier transform is a simple and effective time-frequency analysis tool, but its performance is largely affected by window function, window length and window sliding step size. Long windows can provide better frequency resolution but poorer time resolution, and vice versa. At the same time, window function and sliding step size can also have influence on the time-frequency analysis of the signal. For better time-frequency representation We present an adaptive method based on reinforcement learning, which can adaptively and synchronously adjust three parameters according to different data characteristics. Simulation results show the adaptive method can dramatically increase the time-frequency resolution of short-time Fourier transform.