Abstract

This paper proposes a new clustering method called Gustafson-Kessel with Focal Point (GKFP). The proposal aims at benefiting from the advantage of using Gustafson-Kessel clustering technique leveraged by the use of a Focal Point which enables obtaining partitions with different levels of granularity. Thus the method identifies clusters with uncorrelated or strongly correlated data while it allows the user to explore different regions of the feature space with different levels of detail. Due to the possibility of dealing with correlated data, a regularization procedure might be necessary. Therefore, the paper also briefly describes a Bayesian regularization which can be associated with GKFP. Experiments from bearing fault diagnosis show that GKFP outperforms three other clustering techniques, i.e., the popular fuzzy c-means (FCM), Gustafson-Kessel (GK), and the state of the art FCMFP, for two different bearing data sets.