Signal classification via multi-scale PCA and empirical classification methods

Abstract

Data coming from a real–world complex system are usually contaminated by noises or some irrelevant components, which do not contribute to improve signal classification accuracy. Also in the process of signal feature enhancement, the performance of any statistical method used to recover the original signals may be impacted by the noise. In this paper, we propose the multi–scale principal component analysis (PCA) method, which combines discrete wavelet transform with PCA for feature enhancement and signal decomposition in both spatial and temporal domains. We developed a new classification method, called empirical classification (EC), to classify the power spectra of the feature extracted signals after the multi–scale PCA procedure. These methods were applied to a publicly available EEG database for the purpose of signal classification. An overall accuracy of 99% for the classification of 500 real EEG recordings under different considered classification problems is obtained. Our results show that signal decomposition by multi–scale PCA coupled with the EC method, leads to a highly promising accuracy in classifying epileptic EEG signals.