Generating Controlled Physics-Informed Time-to-failure Trajectories for Prognostics in Unseen Operational Conditions

The performance of deep learning (DL)-based methods for predicting remaining useful life (RUL) may be limited in practice due to the scarcity of representative time-to-failure (TTF) data. To overcome this challenge, generating physically plausible synthetic data is a promising approach. In this study, a novel hybrid framework is proposed that combines a controlled physics-informed data generation approach with a DL-based prediction model for prognostics. The framework introduces a new controlled physics-informed generative adversarial network (CPI-GAN) that gen-erates diverse and physically interpretable synthetic degradation trajectories. The generator includes five basic phys-ics constraints that serve as controllable settings. The regularization term, which is a physics-informed loss function with a penalty, ensures that the synthetic data's changing health state trend complies with the underlying physical laws. The synthetic data is then fed to the DL-based prediction model to estimate RUL. The framework's effective-ness is evaluated using the New Commercial Modular Aero-Propulsion System Simulation (N-CMAPSS), a turbo-fan engine prognostics dataset with limited TTF trajectories. The experimental results demonstrate that the proposed framework can generate synthetic TTF trajectories that are consistent with underlying degradation trends and sig-nificantly improve RUL prediction accuracy.

Keywords: Prognostics, Time-to-failure trajectory generation, Deep learning, Physics-informed generative adver-sarial networks.

Download PDF