Remaining Useful Life Prediction Based on Transformer with A Tiny Representation Network

Abstract

Remaining useful life (RUL) prediction is of great significance to the prognostic and health management of rolling bearings. The effectiveness of the typical RUL prediction relies on the constructed health indicator (HI) which only represents limited degradation information. In addition, rolling bearing degradation is a long-term process, while existing RUL prediction models show a limited ability to learn a long-distance dependency. To fill the above research gap, we propose a novel RUL prediction Transformer (RPT) which consists of a tiny convolution-based representation network (RN) and an advanced Transformer feature extractor. In the proposed RPT, the row vibration signals are concisely and efficiently embedded into a tiny feature space by the RN. Then, embedded vectors of historical run-to-failure data are input into the transformer feature extractor to learn potential prediction knowledge. Due to the global attention machine, the RPT can learn long-distance dependency, which significantly improves the RUL prediction. Compared with state-of-the-art models, RPT attains more accurate RUL prediction.