Multi-biophysical event detection using blind source separated audio signals

Abstract

This paper aims to use signal processing techniques to identify biophysical events using audio signals. The processing technique proposed is a combination of the mel frequency cepstral coefficients (MFCC) used as features, independent component analysis (ICA) and principle component analysis (PCA) algorithms to separate sources and noise. It was found that compressing the data into the energies of 26 filter banks mapped to the mel frequencies, sufficient descriptive information was conserved as validated by visually identifiable source signal patterns. Subsequently performing PCA isolated global background noise to an individual component. Further, by performing ICA, components contained independent and visually identifiable patterns that correlated to events associated with heart rate, squatting motion and involuntary abdominal movements. This componentized feature space provides an optimized source for building a discriminant function for the classifier used in the machine learning algorithm to provide simultaneous and automatic classification of these biophysical events.