In consideration of the problem in applying Kalman filter when time varying noise covariance matric of measurement and process are neither known, a new adaptive filter was purposed by Variational Bayesian(VB) approach. This filter overcame two key problems as following: first, relative transfer probability was proposed as realizing to promote the performances of steady state precision and dynamic respond simultaneously, by designing the heuristic adaptive window for noise estimation, according to the conjugate posterior distribution of measurement and process noise; second, the serious offset of the estimation value of noise covariance matric under single time scale was notable reduced by designing the approach which estimated the covariance matric under double time scales. The simulations proved that this method could track the noise statistics feature quickly without losing estimation precision.