Self-supervised Learning for Remaining Useful Life Prediction of Composite Materials based on Physics-based Synthetic Strain Data

In the present research, a self-supervised learning (SSL) framework is proposed for predicting the remaining useful life (RUL) of composite laminates using physics-based synthetic strain data generated from high-cycle fatigue simulations. Although deep learning (DL) models have been increasingly applied to prognostic problems, the limited labeled data remain a major challenge for effective model training. To overcome data scarcity, an SSL approach, an emerging DL technique, is proposed in this study. The advantage of SSL is that it leverages a large unlabeled dataset (i.e., containing only raw strain gauge data) while requiring only a small labeled dataset (i.e., containing both strain and RUL data). The datasets are generated through simulations, in which the delamination propagates with fatigue cycles, and strain data are recorded using virtual strain gauges. By sampling fatigue parameters from statistical distributions, a large-scale dataset can be generated from a single simulated structure. The proposed SSL model comprises a pre-training stage on unlabeled data followed by fine-tuning on labeled samples. The influence of pre-training dataset size and the number of labeled structures is evaluated, and the RUL prediction accuracy of the proposed time-aware self-supervised framework is validated. The results demonstrate the feasibility of the proposed time-aware SSL framework for RUL prediction within the synthetic DCB delamination dataset.