Statistical neighbor classification (SNC) for condition monitoring of rotating machines

Variation-based Condition Monitoring (VCM) is one of the most com-mon techniques for detecting faults on rotating machines. This is in continuous development and this study approaches the faults detection through pattern recog-nition concepts in order to respond to industry demands for automated fault detec-tion systems. For this purpose, a set of five common shaft-related conditions in-cluding healthy, unbalance, misalignment, looseness and rub have been experimentally simulated at two rotational speeds. The study has been designed to use common VCM techniques such as Time Domain Analysis (TDA) and Fre-quency Domain Analysis (FDA), which are well known within industry. As this study forms part of ongoing studies on automatic faults classification at the Dy-namics Laboratory of the University of Manchester, Principle Component Analy-sis (PCA) that was used in an earlier exploratory study was similarly investigated on a combination of TDA and FDA parameters. This study shows that PCA can be ineffective when a set of measurements representing machine conditions with high variability of fault features but may lead to cluster overlap when faults are very similar in severities. Based on this premise, the Statistical Neighbor Classification (SNC) approach, inspired by the k-Nearest Neighbor (kNN) concept was ex-plored for refinement and the preliminary results presented here shows that SNC possesses the potentials to compensate PCA when dealing with faults with very similar severities.