Spatial-Temporal Feature Extraction Network for Online Aeroengines Remaining Useful Life Prediction

Abstract

Remaining useful life (RUL) prediction based on health information is of vital significance in prognostics and health management (PHM). In recent years, accuracy RUL prediction by data-driven methods is the hotpot. However, most data-driven RUL prediction methods lack the ability to capture spatial-temporal features from input monitoring data, which makes carrying out RUL prediction difficult under online scenarios. To overcome this weakness, an online RUL prediction framework is proposed with spatial attention temporal convolutional network (SATCN). Specifically, the spatial attention (SA) module can initially capture features’ properties to quantify the intrinsic spatial structure, which is the internal spatial feature. The external spatial feature can be acquired by calculating the spatial relationship value between different features. Then, the temporal convolutional network (TCN) is constructed by causal convolution and dilated convolution to extract the temporal feature. Based on the SATCN module, an online RUL prediction framework is proposed. In this framework, a new window time input data and an online loss function are defined to ensure that model parameters can be updated online. Also, new extended data are added to the current input, which can improve the RUL prediction accuracy. Finally, the obtained results from aeroengine datasets demonstrate the superiority of the SATCN-based online RUL prediction framework compared to existing data-driven methods.