A novel feature-fusion-based end-to-end approach for remaining useful life prediction

Remaining useful life (RUL) prediction is a key aspect of health condition monitoring, which can reduce maintenance costs and improve system operational efficiency. The most existing approaches only extract temporal features or spatial features, and ignore raw mapping features in RUL prediction. However, these different features are highly complementary and relevant for RUL prediction. Different from these approaches, we propose a novel feature-fusion-based end-to-end approach for RUL prediction in this paper, which combines spatiotemporal features and raw mapping features. To begin with, the time attention mechanism is used for the input to weight different time steps. Then convolutional neural networks (CNNs) are used for the weighted input to extract spatial feature maps. Between the CNNs, channel attention and spatial attention mechanisms are applied to the feature maps to learn the importance of channel and spatial distribution. Meanwhile, a bidirectional gated recurrent unit is adopted to capture temporal dependency features. In addition, the raw mapping features are obtained from the input through a fully connected layer to provide additional information. Finally, the three types of obtained features are fused for the final RUL prediction through fully connected networks. Extensive experiments are carried out on the C-MAPSS engine dataset. The results show that the proposed approach outperforms the current deep learning approaches.