An unsupervised and enhanced deep belief network for bearing performance degradation assessment

Abstract

Abstract An improved unsupervised deep belief network (DBN), namely median filtering deep belief network (MFDBN) model is proposed in this paper through median filtering (MF) for bearing performance degradation. MFDBN has the following advantages: (1) MFDBN uses the absolute amplitude of the original vibration signal as direct input to extract HI and reduce dependence on manual experience. (2) Unlike the bearing failure signal, the degradation signal is continuously changing; hence it is difficult to label the data. To solve this problem, MFDBN is used to extract health indicator (HI) without an output layer. (3) Because the noise of the vibration data, the smoothness of the extracted HI is poor, it is easy to misjudge the bearing health status. The multiple hidden layers with the MF model can denoise the HI curve layer by layer. Finally, this is verified by comparing other models and using multiple bearing datasets to demonstrate its superiority.