Mean shift based segmentation for time frequency analysis of packet based radio signals

We consider the problem of RF signal analysis where one sensing node observes a frequency band possibly used by multiple packet based radio transmitters producing signals with non-persistent excitation. Since each combination of active transmitted signals results in a composite signal with its distinct statistical properties, the received signal at the sensing node consists of a number of statistically homogeneous segments. First important task in the analysis of this type of signals is to localize these segments in time. We propose a segmentation algorithm for solving this problem. Initial segmentation is obtained using a variant of mean shift algorithm with adaptive scale parameters. We provide a convergence analysis of this algorithm and propose a method for selecting scale parameters. Final segmentation results are obtained after removal of impulse noise from the initial segmentation results. Proposed algorithm is almost completely nonparametric and it finds the number of segments automatically. Performance of the algorithm is studied using simulations involving signals used in 802.11 networks. The algorithm can be applied in a scenario where the sensing node is part of a radio scene analysis network providing information that can be used for achieving efficient utilization of radio spectrum.