Dilated attention ConvGRU for remaining useful life prediction of rolling bearings with uncertainty quantification

Abstract In the current industrial field, rolling bearings have become fundamental components of numerous mechanical systems. When the remaining useful life (RUL) of rolling bearings is completed, it is very important to carry out the industrial production process. However, it is common for the characteristics of early small defects in rolling bearings to be difficult to identify, and not obvious. To enhance the accuracy of RUL prediction, we propose a gated neural network framework that combines a dilated 1D convolutional gated recurrent unit with a temporal attention mechanism. We use Monte Carlo Dropout to generate uncertainty intervals. Two datasets were chosen for the experiments: XJTU-SY and PHM2012 dataset. As indicated by the experimental results, our model achieved good performance in both cases, which validates its practical engineering value.