The instant invention relates to a method for noise reduction from a magnetic resonance sounding (MRS) oscillating signal, and more particularly, to a data processing method for reducing random noise contained in MRS oscillating signal based on joint algorithm principles of EMD and TFPF. A MRS oscillating signal is decomposed into different eigen-mode components by using decomposition characteristic of EMD algorithm; then a signal-dominated eigen-mode component is encoded as an instantaneous frequency of an analytical signal of unit amplitude using TFPF algorithm; and random noise is suppressed with the characteristics that the time-frequency distribution of the analytical signal is concentrated along with the instantaneous frequency. The method requires fewer filtering constraints and is simple to operate without need of designing a filtering interval in the time-frequency domain, and has good adaptability to the MRS oscillating signal with a low signal-noise-ratio.