Matching pursuit classification for time-varying acoustic emissions

Time-varying techniques have been shown to be successful for classifying acoustic signals. In this paper, we propose a time-frequency based method for classifying time-varying, transient acoustic emissions using the matching pursuit decomposition (MPD) algorithm. Before classification, the MPD dictionary is formed using time and frequency shifted versions of selected learning signals. We show the superior classification performance of this method by applying it to real data obtained from an acoustic monitoring system. The main objective of an acoustic signal classifier in this system is to identify the acoustic emissions that may occur in concrete structures in distress. We study different strategies for binary and M-class classification, their dependence on the learning signals, and our classification results. We also discuss techniques to reduce the MPD processing time by pre-processing the test signals.