Bilinear time-frequency distribution can token overall mechanical failure signal characteristics, but it's found that strong cross-terms exist which results in frequency aliasing and information loss appear. In this paper, the time-frequency analyze is used to identificate the weak winking signal in complex background. Arithmetic based on ARMA model filter is bring forward to solve cross-terms problem. The arithmetic is simulated in experiment data and contrasted to Smooth- Puppet Wigner-Ville arithmetic. The conclusion is that arithmetic of ARMA model pre-filter restrained cross-terms disturbance better and is of better weak winking signal identification ability.