Hierarchical attention graph convolutional network to fuse multi-sensor signals for remaining useful life prediction

Highlights: • A novel multi-sensor features fusion approach for remaining useful life prediction is proposed;. • Multiple sensors are constructed to a sensor network to generate spatial-temporal graphs. • Hierarchical attention graph convolutional network is constructed to model the generated spatialtemporal graphs. • Regularized self-attention graph pooling is proposed to obtain more informative graph representations. Deep learning-based prognostic methods have achieved great success in remaining useful life (RUL) prediction, since degradation information of machine can be adequately mined by deep learning techniques. However, these methods suffer from following weaknesses, that is, 1) interactions among multiple sensors are not explicitly considered; 2) they are more inclined to model temporal dependencies while ignoring spatial dependencies of sensors. To address those weaknesses, the multiple sensors are constructed to a sensor network and hierarchical attention graph convolutional network (HAGCN) is proposed in this paper for modeling the sensor network. In HAGCN, the hierarchical graph representation layer is proposed for modeling spatial dependencies of sensors and bi-directional long short-term memory network is used for modeling temporal dependencies of sensor measurements. Moreover, a regularized self-attention graph pooling is designed in HAGCN to achieve effective information fusion of the sensors. To realize prognostics, the spatial-temporal graphs are firstly generated based on the sensor network. Then, HAGCN is applied to model the spatial and temporal dependencies of the graphs simultaneously. The experimental results of two case studies show the superiority of HAGCN over state-of-the-art methods for RUL prediction.