A vision subsampling probsparse self-attention-based informer for bearing remaining useful life prediction

Abstract

Bearing remaining useful life (RUL) prediction has always been a central topic in the industry field, the aim of which is to optimize system safety and sustainability. The validity of prediction models and the accuracy of prediction results are affected by mid-term singularities and terminal mutations, under time-domain bearing vibration information. In this paper, a network structure-cascaded dilated convolution vision informer (CDC-Vii) is put forward to precisely forecast the RUL of bearings, which uses the time–frequency fault features as input. CDC-Vii breaks the limitation of the original Informer, which is only sensitive to time-series information. An adaptive fault frequency band selection algorithm is proposed, which can reduce training time while utilizing rich time–frequency information. Based on the Informer architecture, the attention mechanism is improved to form vision subsampling probsparse self-attention (VSPS). VSPS can precisely assign spatial attention weights and reduce computational complexity. At the same time, a truncated relative position encoding technique is proposed to strengthen the position dependence between attention information. Moreover, cascaded dilated convolution enhances the image contrast of faulty frequency bands while enlarging the use of the receptive field. Experiments on two extensively utilized bearing datasets reveal that CDC-Vii surpasses the advanced RUL prediction models.