A spatiotemporal fused bidirectional temporal convolutional network for remaining useful life prediction

Abstract To guarantee the dependable functioning of complex mechanical systems, remaining useful life (RUL) prediction has emerged as a pivotal research frontier. Temporal convolutional networks (TCNs) have revolutionized RUL prediction through innovative integrations of convolutional neural networks and attention mechanisms. However, these integrations have not paid significant attention to the fusion of spatiotemporal information. Furthermore, conventional TCNs fail to bidirectionally exploit the sequence information to predict RUL. Here, we propose a spatiotemporal fused bidirectional temporal convolutional network (STF-BiTCN) for predicting RUL. First, a long-short term attention unit is designed to enhance feature representations in temporal dimension. Then, a multi-dimensional convolution unit is designed, which allows for dynamic adjustment of the convolutional kernel to fully capture the spatial features. Finally, the BiTCN is devised to bidirectionally exploit the sequence information of spatiotemporal fused features, which contributes to sufficiently learning the correlations between degradation features and RUL labels. Several experiments are conducted on two individual datasets to verify the predictive accuracy of the developed model.