Automated novelty detection using a modified Kohonen self organizing map

This paper proposes a novelty detection based method for machine condition monitoring (MCM) and a new feature extraction method based on higher order statistics of the power spectral density. This novel MCM method is structured on a modified Kohonen's self-organising map. It adopts a multi-dimensional dissimilarity measure for dual-class classification. The approach is designed to be highly modular and fits in well in a multi-sensor condition monitoring environment. Experiments using real world datasets with up to eight sensors have shown high accuracy in classification and robustness across different CM applications.