Novelty Detection in Machine Vibration Data Based on Cluster Intraset Distance

Traditional methods for condition monitoring of machinery are based on detecting situations when values of features extracted form measurement data leave predetermined bands consistent with normal machine operation. Design of such systems requires considerable amounts of measurement data describing machinery failure modes that are generally very difficult to obtain due to large number of failure examples necessary. Novelty detection is the identification of new or unknown data or signal that a machine learning system is not aware of during training. Assumption is that in case of impeding failure, new previously unseen measurement data will appear. In this paper method of novelty detection in machine vibration data is described based on clustering of features extracted from measurement data. During training system discover main operational regimes of machine and assign to them clusters of feature data. Later, during machine exploitation, by comparing intraset distances within cluster members with closest distance of new example to cluster centers, system is able to detect abnormal new measurement data that were not known at the time of training the model. Method is presented on car engine vibration data.