Trend attention fully convolutional network for remaining useful life estimation

Modern engineered systems usually employ multiple sensors to monitor equipment health status. However, most remaining useful life (RUL) estimation methods based on deep learning are hard to select helpful signals and remove useless signals accurately. Moreover, the attention mechanisms they employed could hardly obtain an optimal attention distribution at an acceptable computational cost, resulting in poor prediction performance. Therefore, we proposed a novel signal selection method, terming the "Loss boundary to Mapping ability" (LM) approach. It can accurately select the signals that can contribute to RUL prediction tasks. Then, inspired by the characteristics of RUL monitoring signals, we proposed a novel end-to-end framework called Trend attention Fully Convolutional Network (TaFCN) to enhance prediction performance further. These two methods constitute our prognostic method. We conducted a series of ablation experiments and comparative experiments with recent methods on the C-MAPSS turbofan engine dataset. The ablation experiments proved the necessity and advanced performance of the LM and the proposed attention mechanism employed in the TaFCN. The comparative experiments demonstrated the state-of-the-art performance of our prognostic method. Furthermore, we developed an interpretability analysis method, which revealed the logical reasoning process of our method.