Fault diagnosis of rotor systems using ICA based feature extraction

Abstract

A method is proposed for fault diagnosis of rotor systems, with independent component analysis (ICA) based feature extraction and multi-layer perceptron (MLP) based pattern classification. By the use of ICA, feature vectors are integratedly extracted from multichannel vibration measurements collected under different operating patterns (in term of rotating speed and/or load). Thus, a robust multi-MLP classifier insensitive to the change of operation conditions is constructed. Experimental results indicate invariable fault features embedded in vibration observations can be effectively captured and different fault patterns (for example imbalance, impact and loose foundation) can be correctly classified, both of which imply great potential of the proposed ICA-MLP classifier in fault diagnosis of rotor systems.