Memory-enhanced hybrid deep learning networks for remaining useful life prognostics of mechanical equipment

Abstract Remaining useful life (RUL) prognostics is one of the most important parts in prognostics and health management, which can effectively avoid sudden accidents and economic losses caused by mechanical equipment failure. Previous artificial intelligence based prognostics method depends on manual feature extraction, which requires a lot of expert experience and prior knowledge for feature design. This paper presents a memory-enhanced hybrid deep learning network (MEHDLN) combining convolution neural network and recurrent neural network, which contains convolution layer, pooling layer, bidirectional long short-term memory (BLSTM) layer and fully connected (FC) layer. It can not only extract local robust features from raw signals but also capture time-dependency in sequence sensor signals. The proposed MEHDLN model is verified by two experimental cases including rolling element bearing and turbine engine RUL prediction. Experimental results demonstrate the superiority of the MEHDLN over other state-of-the-arts.