Empirical mode decomposition based sparse dictionary learning with application to signal classification

This paper will present a novel empirical framework for dictionary learning where the dictionary is learned from the data to be analyzed, rather than using a pre-defined basis. A dictionary formation and learning algorithm is presented, which learns sparse dictionaries, where sparsity is understood in terms of the small number of dictionary atoms compared to the signal dimensions. An initial dictionary is formed using training signals of different classes, where the dictionary atoms consist of intrinsic mode functions obtained as a result of decomposing the training signals using empirical mode decomposition. A dictionary learning algorithm trains this dictionary which results in a significant reduction in the size of the learned dictionary. The learned dictionary can be applied to signal classification, whereby coefficients of orthogonal projections of test signals against the learned dictionary are used as features to classify the test signals into different classes. We also show that the learned dictionary allows calculation of the coefficient vector based on sparse representation of test signals, which can also be used as a feature vector. Although the framework is not formulated as reconstructive, or combined reconstructive and discriminative dictionary learning, its efficacy in signal classification is demonstrated using real-life EEG signals.