Industrial acoustic signal processing with graph based features

Abstract

Although acoustical features can be extracted directly from time series, more relevant and more precise features can be collected from a higher processing level, namely, the frequency domain. Physicians prefer thin peaks in the frequency space, which can be usually achieved by windowing and wavelet analysis. In a bio-inspirited way, the human brain can determine an object from the area of the region enclosed by the curve of a function. In this paper a higher process level is demostrated where a graph-based feature extraction algorithms is used on the Auto Power Spectrum Density (APSD) function of an acoustical signal. There are three main approaches to calculate the Medial Axis of this binary object: thinning, distance-based skeletonization and Voronoi skeletonization. We found that the latter one serves best our purpose, because it uses the least points to generate a tree graph. This tree can be analysed by dataminer algorithms which are well-known in the field of machine learning, thus the resulting structure serves as input for classification methods.