Sequence-to-Sequence Remaining Useful Life Prediction of the Highly Maneuverable Unmanned Aerial Vehicle: A Multilevel Fusion Transformer Network Solution

Abstract

The remaining useful life (RUL) of the unmanned aerial vehicle (UAV) is primarily determined by the discharge state of the lithium-polymer battery and the expected flight maneuver. It needs to be accurately predicted to measure the UAV’s capacity to perform future missions. However, the existing works usually provide a one-step prediction based on a single feature, which cannot meet the reliability requirements. This paper provides a multilevel fusion transformer-network-based sequence-to-sequence model to predict the RUL of the highly maneuverable UAV. The end-to-end method is improved by introducing the external factor attention and multi-scale feature mining mechanism. Simulation experiments are conducted based on a high-fidelity quad-rotor UAV electric propulsion model. The proposed method can rapidly predict more precisely than the state-of-the-art. It can predict the future RUL sequence by four-times the observation length (32 s) with a precision of 83% within 60 ms.