Abstract：An intelligent fault diagnosis model based on individual feature selection (IFS) and relevance vector machine (RVM) is proposed for bearing fault diagnosis under varying load conditions. First, vibration data measured under no-load (0hp) and full-load (3hp) is combined and used for training. Second, Statistical characteristics in time domain and node energies in full wavelet packet domain are extracted as candidate features. Third, an improved Fisher feature selection method is proposed and used to select individual best feature subset for each pair of classes. Fourth, using one-against-one (OAO) approach, several RVM binary classifiers are trained for each pair of classes. Finally, the IFS_RVM multi-class fault diagnosis model is constructed by combining all RVM binary classifiers using max-probability-win (MPW) strategy. Vibration data measured under untrained load conditions (1hp and 2hp) are used for testing, resulting in very high accuracy of 99.58%. Experimental results demonstrate that the proposed model is very effective and robust for online fault diagnosis under varying load conditions.
2012, Vol. 31 
