Background spectrum estimation via robust Kalman filtering

This study deals with the passive SONAR signal background spectral estimation. Practically, this processing is necessary to detect acoustic vibration from the data gathered by that kind of device. These phenomena appear as peaks on the estimated spectra of the collected data. That's why a decision test is applied on the estimated power spectral density. In order to ensure a constant false alarm rate of the detector, one needs to normalize the spectra, i.e. split each spectrum into three parts: the peaks, the background and a superimposed noise. Among the whole different technique developed during the last decades, the processing presented in this paper is the robust Kalman filter, an EM processing where the E step is a Kalman filter step and the M step is a dynamical system parameters estimation. This framework presents the interest to be real time and full automated, and not signal dependent, as long as the system initial guess remains physically realistic. Experimentations on simulated data and real world data are presented.