Riemannian Distances for Signal Classification by Power Spectral Density

Signal classification is an important issue in many branches of science and engineering. In signal classification, a feature of the signals is often selected for similarity comparison. A distance metric must then be established to measure the dissimilarities between different signal features. Due to the natural characteristics of dynamic systems, the power spectral density (PSD) of a signal is often used as a feature to facilitate classification. We reason in this paper that PSD matrices have structural constraints and that they describe a manifold in the signal space. Thus, instead of the widely used Euclidean distance (ED), a more appropriate measure is the Riemannian distance (RD) on the manifold. Here, we develop closed-form expressions of the RD between two PSD matrices on the manifold and study some of the properties. We further show how an optimum weighting matrix can be developed for the application of RD to signal classification. These new distance measures are then applied to the classification of electroencephalogram (EEG) signals for the determination of sleep states and the results are highly encouraging.