Data Driven Damage Evaluation and Prognostics in Aerospace Composites: Integration of Experiments, Simulations and Machine Learning

This study presents an integrated structural health monitoring (SHM) framework applied to composite structures, combining experimental and numerical analyses with supervised, unsupervised, and reinforcement learning (RL) techniques. The investigated structure is a carbon-epoxy composite laminate [(45/-45/0/90)4]s , representative of aerospace applications. Experimental tests were conducted using surface-mounted piezoelectric (PZT) sensors to excite and capture guided Lamb wave signals and vibration responses. Parallel to the experiments, a finite element (FE) model was developed to simulate guided wave propagation and dynamic behavior, enabling validation and parametric exploration of damage scenarios. Signal features such as time-of-flight, peak amplitude, Root Mean Square (RMS) energy, spectral centroid, and cross-correlation were extracted. A correlation analysis was performed to identify the most informative features for training (ML) machine learning models. Three approaches were implemented and compared: Random Forest (supervised), K-Means (unsupervised), and Q-learning (reinforcement). The models aimed to classify damage severity, detect anomalies, and estimate remaining useful life (RUL). The combined experimentalnumerical approach demonstrated good agreement between observed and simulated waveforms. Results show that supervised learning achieved the highest damage classification accuracy, while RL provided better adaptation to unseen scenarios. Feature correlation analysis also revealed key indicators for reliable prognostics. This study contributes a comparative framework for SHM implementation in composite structures and supports the transition toward predictive maintenance in aerospace applications.