Mazzoleni M., S. Formentin, F. Previdi, S.M. Savaresi
In this paper, the use of the Principal Direction Divisive Partitioning (PDDP) method for unsupervised learning is discussed and analyzed with a focus on fault detection applications. Specifically, a geometric limit of the standard algorithm is highlighted by means of a simulation example and a modified version of PDDP is introduced. Such a method is shown to correctly perform data clustering also when the standard algorithm fails. The modified strategy is based on the use of a Chi-squared statistical test and offers more guarantees in terms of detection of a wrong functioning of the system. The proposed algorithm is finally experimentally tested on a fault detection application for aerospace electro-mechanical actuators, for which a comparison with k-means and fuzzy k-means approaches is also provided.
Download
