Second order notch lters are commonly used for frequency estimation purposes. They are characterized by a single design parameter, the value of which is selected to achieve a suitable compromise between tracking performance and bias/variance errors. We study the problem of estimating the dominant frequency of a non stationary Gaussian series by using notch lters. The design parameter is performed using wavelet tool, the technique thresholds the empirical wavelet coecients in order to estimate the bandwidth. The classical Recursive Least Square Algorithm runs on the wavelet packet tree choosing the more sensitive time frequency level. We propose an empirical on-line algorithm using wavelet packets for detection of pantograph vibration in rail vehicle.