Prediction of Remaining Service Life of Rolling Bearings Based on Convolutional and Bidirectional Long- and Short-Term Memory Neural Networks

Predicting the remaining useful life (RUL) of a bearing can prevent sudden downtime of rotating machinery, thereby improving economic efficiency and protecting human safety. Two important steps in RUL prediction are the construction of a health indicator (HI) and the prediction of life. Traditional methods simply use the time-series characteristics of the vibration signal, for example, using root mean square (RMS) as HI, but this HI does not reflect the true degradation of the bearing. Meanwhile, existing prediction models often cannot consider both the time and space characteristics of the signal, thus limiting prediction accuracy. To address the above problems, in this study, wavelet packet transform (DWPT) and kernel principal component analysis (KPCA) were combined to extract HI from the original vibration signal. Then, a CNN-BiLSTM (convolutional and bidirectional long- and short-term memory) prediction network with root mean square as input and HI as output was constructed by combining convolutional neural network (CNN) and bi-directional long- and short-term memory neural network (BiLSTM). The network improved prediction accuracy by considering the temporal and spatial characteristics of the input signal. Experimental results on the PHM2012 dataset showed that the method proposed in this paper outperformed existing methods.