Remaining useful life prediction of bearings based on temporal convolutional networks with residual separable blocks

It is essential for the remaining useful life (RUL) prediction of bearings to ensure the safe operation of rotating machinery. Rotating machines are highly complex. However, the critical degradation information of bearings is often neglected due to the insufficient perceptual field of temporal convolutional networks, which results in poor prediction results. To solve the problems, a new framework named temporal convolutional network with residual separable convolutional block (TCN-RSCB) is proposed in this paper for bearing RUL prediction. First, RSCB is constructed by using residual learning and separable convolution. In RSCB, in order to obtain a larger perceptual field, we continuously increase the dilation factor of the separable convolution according to the exponential level. Then, a soft thresholding temporal convolution block (STCB) is constructed by using a soft thresholding technique, this block can eliminate the redundant information in the prediction network and improve the prediction results of bearing RUL. Finally, the proposed method is tested on the FEMTO-ST dataset and the XITU-SY dataset, and compared with some advanced prediction methods. The results indicate that TCN-RSCB can follow the real RUL well on different bearing datasets, and has good robustness and generalization ability.