Underwater target radiated noise signal possesses the features of non-stationary, non-Gaussian and strong background noise. It's difficult to detect characteristics at low signal-noise ratio. Empirical Mode decomposition algorithm is used to handle tranquilization of radiation noises, and then filtered sub-band signal is fed into the improved model of stochastic resonance. By changing internal noise intensity of the system, enhancement of the weak periodic signal is realized under the synergies of system, signal and noise. The simulation results show that the weak signal power spectrum value is improved nearly 25 dB when the algorithm is used for detecting characteristics of actual underwater signal.