Fault detection via modified Principal Direction Divisive Partitioning and application to aerospace electro-mechanical actuators

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.