Bearing fault diagnosis based on improved ensemble learning and deep belief network

Abstract

The collected bearing signals are easily interfered by strong ambient noise due to complex operating conditions. It’s a challenge to identify faults accurately and to reduce dependence on model hyper-parameters for intelligent diagnostic methods. This paper proposes an improved ensemble method based on deep belief network (DBN) for fault diagnosis of rolling bearings. Firstly, a series of DBNs with different hyper-parameters are constructed and trained. Secondly, the improved ensemble method is used to acquire the weights matrix for each DBN. Finally, each DBN votes together in accordance with its respective weight matrix to get the final diagnosis result. This method is applied to rolling bearing data from Case Western Reserve University. Experiments show that the effect of fault diagnosis is significantly improved because the feature learning ability of different DBNs is fully taken advantage of.