A study on remaining useful life prediction methods based on physics-informed neural networks

The process of estimating system time serves as vital component for upkeep operations within industrial facilities that contain turbofan engines, which represent intricate equipment. Traditional machine learning (ML) algorithms face difficulties when encountering such challenges as processing sparse data, handling noisy sensor measurements, and modeling complex system dynamics. The research seeks to improve useful life (RUL) forecasting precision through the implementation of Physics-Informed Neural Networks combined with deep learning (DL) technologies. The development of a model to predict RUL must achieve two requirements, including precise forecasting of all RUL values and ability to generate predictions consistent with physical degradation patterns. This research introduces a hybrid PINN-based system that combines data-driven learning with physics-based constraints. The NASA C-MAPSS dataset serves as validation dataset for study. Experimental results demonstrate that the proposed model outperforms traditional ML algorithms and existing DL models. The model achieves a Root Mean Squared Error of 0.02184, a Mean Squared Error of 1.0000, and a coefficient of determination value of 0.953. The study presents three key contributions: a hybrid RUL prediction framework integrating attention-based feature extraction with PINN, the DeepHPM dynamics module incorporating degradation constraints, and improved predictive accuracy on the FD001 dataset.