Combination of SGDD and Multiple Time Scale CNN for Bearing Fault Diagnosis

Abstract

This paper aims at the fault identification of common types of the rolling element bearings (REBs). A method employs a Savitzky-Golay Digital Differentiator (SGDD) and convolutional neural network (CNN) with multiple time scale mechanism is proposed. The SGDD produces the derivative of the estimated smoothed signal, which can present jerk, the derivative of acceleration. And jerk is a better alternative to extract features from, whose limited usage is mainly due to the high price. The jerk signal is input to ID CNN with modified structure, namely the down-sampling of input signal and the convolution layers and pooling layers for signals of different time scales respectively. Their output are then concatenated. The remaining procedures are identical to them of an essential CNN. The method is validated with a dataset of REB. The results indicate that the approach has a higher accuracy and runs more smoothly during training process than other two related methods, thus confirm the role of SGDD and multiple time scale mechanism in REB fault classification.