 |
International Society of Science and Applied Technologies |
| RUL Prediction of Aeroengine Gas Path Systems Using Gradient-Enhanced Trajectory-Parameter Ensemble Learning | ||||
| Author |
Jiaxuan Zhan
|
|||
| Co-Author(s) |
Fanping Wei; Xiaobing Ma
|
|||
| Abstract | Accurate prediction of remaining useful life (RUL) for critical equipment such as aero-engines is the core of predictive maintenance. However, existing mainstream methods encounter severe challenges. On the one hand, pure data-driven models are capable of learning complex patterns yet heavily depend on sufficient high-quality training data. On the other hand, degradation models based on stochastic processes adopt predefined fixed degradation trajectories, which can hardly adapt flexibly to the complex nonlinear degradation dynamics learned from data, thereby limiting their applications in variable operating environments. To effectively address the above dual challenges, this paper proposes a physics-informed enhanced adaptive trajectory learning method. A hybrid framework is established to synergize the strengths of data-driven methods and physical models. It leverages the global pattern learning ability of deep learning to capture complex nonlinear degradation trends, and embeds basic physical laws as hard constraints to ensure the physical consistency and stability of RUL prediction results. Moreover, an online adaptive mechanism is designed in the framework to rapidly adapt to performance differences among individual equipment units. Experimental results on the benchmark NASA C-MAPSS dataset show that the proposed method outperforms standalone data-driven models and stochastic process models in prediction accuracy.
|
|||
| Keywords | Remaining Useful Life Prediction, Physics-Informed Neural Networks, LSTM, Gamma Stochastic Process | |||
| Article #: RQD2026-241 | ||||
Proceedings of 31st ISSAT International Conference on Reliability & Quality in Design |