VPMCD based novelty detection method on and its application to fault identification for local characteristic-scale decomposition

As a new multi-class discrimination approach, variable prediction model class discrimination (VPMCD) can make full use of the intrinsic relationship among fault features to built variable prediction model for different working conditions and to accomplish multi-class discrimination according to prediction error square sum. It has been effectively used to multi-fault diagnosis when typical fault samples and fault modes can be obtained. However, in most application cases, there only exists normal samples, or there are short of typical fault modes; therefore the variable prediction model is unable to be established and there appear a challenge. Aiming at this problem, VPMCD-based novelty detection (VPMCD-ND) method is put forward in this paper. In VPMCD-ND method, the classifiers are trained only by normal samples firstly. Subsequently, the threshold of prediction error square sum is set according to Chebyshev’s inequality. Lastly, the novelty (from abnormal class) is detected by whether the prediction error square sum is larger than the threshold. Combing with Local characteristic-scale decomposition, a fault diagnosis method is developed and applied to roller bearings. The results show that the proposed VPMCD-ND method not only is more effective than the support vector data description method, but is benefit for online fault diagnosis.