Remaining useful life prediction of roller bearings based on improved 1D-CNN and simple recurrent unit

Abstract

To overcome the shortcomings of traditional roller bearing remaining useful life prediction methods, which mainly focus on prediction accuracy improvement and ignore labor cost and time, the present work proposed a novel prediction method by combining an improved one-dimensional convolution neural network (1D-CNN) and a simple recurrent unit (SRU) network. For feature extraction, the proposed method uses the ability of the 1D-CNN to extract signal features. Moreover, use the global maximum pooling layer to replace the fully connected layer. In the prediction part, a parallel-input SRU network was established by reconstructing the serial operation mode of a traditional recurring neural network (RNN). Finally, experiments were carried out using the XJTU-SY dataset to verify. Results revealed that on the premise of ensuring prediction accuracy, the 1D-CNN-SRU method could reduce manual intervention and time cost to a certain extent and provide an intelligent method for roller bearing remaining useful life prediction.