Multi-attention Based Multi-scale Temporal Convolution Network for Remaining Useful Life Prediction of Rolling Bearings

Remaining useful life (RUL) prediction is the key to prognostic health management (PHM) of rolling bearings. Multiple sensors are usually used to monitor the comprehensive data of the rolling bearing operation. To the best of our knowledge, most of the existing research work does not consider the effects of different sensors and different features to RUL prediction results, so they cannot accurately understand more comprehensive information about rolling bearings. Therefore, it is necessary to weigh the information of different sensors and assign weights to the features extracted by the neural network. This paper proposes a RUL prediction method of rolling bearings based on multi-attention and multi-scale temporal convolution network (MA-MTCN). Specifically, the channel attention mechanism is used to weigh different sensors and the feature attention mechanism is used to weigh different features which extract by multi-scale temporal convolutional networks. The temporal convolutional networks are used to extract the complete features and time series features between data. Finally, the feasibility of this method is verified by PHM2012 dataset. Experimental results prove that the MA-MTCN method can reduce the error of RUL prediction to less than 5%.