A Novel Spatio-Temporal Enhanced Convolutional Autoencoder Network for Unsupervised Health Indicator Construction

The health indicator (HI) of rotating machinery affects the reliability and accuracy of its remaining useful life (RUL) prediction. Convolutional autoencoder (CAE) is widely used for HI construction with the advantage of being unsupervised, but still suffers from two problems: (1) the shallow learning of local spatial features and the neglect of global temporal features result in poor performance of the constructed HI trend; (2) a non-end-to-end form that requires human involvement, which is time-consuming and labour-intensive. Therefore, a novel unsupervised spatio-temporal enhanced convolutional autoencoder (STECAE) network is proposed in this paper, which can directly enhance feature mining on raw data in both temporal and spatial dimensions without any prior knowledge. The STECAE network mainly consists of three components: spatial mining module (SMM), temporal learning module (TLM) and trend compensation mechanism (TCM). SMM enhances the deep mining of local spatial features via deep residual shrinkage module. TLM mines global temporal features through the gate recurrent unit (GRU). TCM acts as a bridge for spatio-temporal connectivity to improve the model's global sensing field and HI's trend performance. The effectiveness of the proposed method is validated through the data set of rolling bearings and aero engines. The results show that compared to the existing state-of-the-art methods, the HI constructed by the proposed method is closer to 1 in terms of the evaluation metric values and has a lower RUL prediction error. The high-quality HI constructed by the proposed method can provide a reliable basis for RUL prediction.