Features extraction for signal classification based on Wigner-Ville distribution and mutual information criterion

The presented method deals with the extraction of features for the classification of non-stationary signals, when the process is only described through training data. The features are determined using the Wigner-Ville distribution (WVD). Three kinds of features are researched: the energy, the temporal expectation and the frequential expectation of the WVD restricted to specific regions. The restriction of the WVD is obtained by applying on the WVD a bidimensional Gaussian window. Given a feature type and a center position of the window in the time-frequency plane, the window parameters are optimized to provide the most discriminant feature. The discriminant nature is measured using a mutual information criterion. This provides a measure of the class separability suitable with any distribution law, and assuming no specific structure of the final classifier. The procedure has been validated with a classification problem of sleep EEG signals.