Abstract: Kernel function approximation is investigated together with some applications in mechanical fault diagnosis, and an approach to rotor fault classification based on feature samples selection is presented. The integral operator kernel functions is used to realize the nonlinear map from the raw feature space of rotor vibration signals to high dimensional feature space, where appropriate feature samples are selected to classify three kinds of rotor faults：rotor crack, rotor unbalance and rotor rub. The quantity of selected samples is much less than that of whole sample sets, which has quickly expedited the computation process. The classification result of KFA is compared with that of SVM. It can be seen that the classification accuracy of KFA is fairly as well as that of SVM, and KFA is or even better than SVM in terms of computation load.

Key words: Fault classification, Feature selection, Kernel function, Rotor, Electromechanical systems, FMEA, Model checking, Reliability evaluation