A multi-fidelity and unsupervised domain adaptation approach for impact detection and localization in composites...

Machine learning techniques have demonstrated strong effectiveness in detecting and localizing low-velocity impacts in composite structures. However, limited experimental data and the high computational cost of high-fidelity (HF) simulations restrict the development of reliable ML models. This study presents a multi-fidelity learning framework that trains Artificial neural network (ANN) using low-fidelity (LF) physics-based simulations, while generalizing to HF experimental data collected from composite flat and stiffened plates. HF signals are noisier and include initial peaks and baseline offsets compared to LF signals, which makes extraction of key features such as Time Difference of Arrival (TDoA) difficult. To address this, a novel TDoA extraction method based on the Bi-Cumulative Signal (BiCS) curve is introduced, enabling consistent TDoA features across fidelity levels. In addition, LF and HF features are aligned using the unsupervised domain adaptation (DA) technique Correlation Alignment (CORAL), with only a small set of unlabelled HF samples. Comparative studies with traditional TDoA methods, and without DA, show that using the proposed BiCS-based TDoA feature together with DA significantly improves impact detection and localization performance. The framework provides a practical and scalable pathway to use ML-based impact monitoring on real composite aircraft structures.