A Time-Frequency Based Method for the Detection and Tracking of Multiple Non-Linearly Modulated Components With Births and Deaths

摘要

The estimation of the components which contain the characteristics of a signal attracts great attention in many real world applications. In this paper, we address the problem of the tracking of multiple signal components over discrete time series. We propose an algorithm to first detect the components from a given time-frequency distribution and then to track them automatically. In the first place, the peaks corresponding to the signal components are detected using the statistical properties of the spectral estimator. Then, an original classifier is proposed to automatically track the detected peaks in order to build components over time. This classifier is based on a total divergence matrix computed from a peak-component divergence matrix that takes account of both amplitude and frequency information. The peak-component pairs are matched automatically from this divergence matrix. We propose a stochastic discrimination rule to decide upon the acceptance of the peak-component pairs. In this way, the algorithm can estimate the number, the amplitude and frequency modulation functions, and the births and the deaths of the components without any limitation on the number of components. The performance of the proposed method, a post-processing of a time-frequency distribution is validated on simulated signals under different parameter sets. The method is also applied to four real-world signals as a proof of its applicability.