Dual residual attention network for remaining useful life prediction of bearings

Abstract

Rolling bearing is a critical component of rotating machines and it is indispensable to accurately predict the remaining useful life (RUL) of bearings to realize predictive maintenance. To extract degradation-sensitive features from complex vibration signals, this paper proposes a new dual residual attention network (DRAN) to improve prediction performance. A frequency band residual attention (FBRA) block is first designed to automatically discover important frequency bands related to bearing degradation. Then, a spatio-temporal residual attention (STRA) block is proposed to sequentially learn high-level representations from frequency and temporal dimensions with a hybrid dilated convolution neural network (HDCNN) and then adaptively identify important features contributing to bearing RUL prediction via a residual attention mechanism. Finally, a weighted RUL estimation block is used to smooth the predicted RUL and provide a more reliable prediction. Experimental results on a public bearing dataset demonstrate the superiority of our DRAN against several state-of-the-art methods.