Classification with automatic detection of unknown classes based on SVM and fuzzy MBF: Application to motor diagnosis
Résumé
Classification algorithms based on data mining tools show good performances for the
automatic diagnosis of systems. However, these performances degrade quickly when the database is
not exhaustive. This happens, for example, when a new class appears. This class could correspond to
a previous unknown fault or to an unknown combination of simultaneous faults. Described algorithm
in this paper proposes a solution to this issue. It combines Support Vector Machine (SVM), fuzzy
membership functions (mbf) and fuzzy information fusion. It results in the construction of a matrix
of memberships to known classes U_class and a vector of membership to unknown classes U_others.
Then, from these values, indicators of distance and ambiguity of the observations can be computed.
These indicators allow setting a simple rejection rule with a threshold classifier. The algorithm is
validated by using Cross-Validation (CV) on experimental data on an induction motor faults supplied
by a voltage-source inverter. The results show the good performances of the proposed algorithms and
its suitability for transportation systems like aircrafts.