Transformer-Based RUL Prediction for Aircraft APU Using Real-World Operational Data

Auxiliary power units (APUs) of aircraft undergo continuous performance degradation during service due to the coupled effects of operating environment and load. Accurate prediction of their remaining useful life (RUL) is of great significance for avoiding unplanned downtime, improving flight safety, and reducing maintenance costs. Most existing APU RUL prediction methods rely solely on a single parameter, which makes it difficult to fully exploit the coupled degradation information among multiple sensors and thus limits prediction accuracy. To this end, this paper proposes a Transformer-based APU RUL prediction model. The model adopts a pure self-attention encoder–decoder structure without any convolutional or recurrent modules, enabling it to automatically extract degradation features directly from real multi-dimensional sensor sequences. It is also free from the constraints of temporal distance within the sequence, offering a notable advantage in modeling long sequences. Experiments on the full-lifecycle real operational data of APUs from an airline demonstrate that the proposed method achieves significantly higher prediction accuracy than mainstream deep-learning baseline models. The proposed method provides a concise and efficient solution for APU RUL prediction.