Feasibility of statistical classifiers for monitoring rollers

In this paper we present our investigations on statistical classification of acoustic signals which is one special assignment in condition monitoring. We compare three pattern recognition methods and three selected feature extraction algorithms with regard to their capability to distinguish between structure-borne sound signatures emitted by intact and worn-out rollers in a drawframe. Our goal is to provide a tool that predicts a forthcoming malfunction of the machine. For this purpose we trained and tested GMM, HMM and SVM based classifiers with spectral, LSF and LCQ features computed from recordings of the operating noise of rollers with varying degrees of abrasion.