Autonomous defect estimation in aluminum plate and prognosis through stochastic process modeling

The structural integrity and longevity of aluminum alloy components in lightweight engineering require accurate and efficient damage detection and prognosis methods. Traditional supervised machine learning (ML) techniques often face limitations due to dependency on large datasets, risk of overfitting, and high computational costs. To overcome these challenges, this study proposes an unsupervised learning framework that combines k-means clustering with a multi-phase gamma process to detect and model damage in aluminum plates. Scanning Acoustic Microscopy (SAM) images serve as the data source, from which comprehensive features are extracted in time, frequency, and time-frequency domains using Short-Time Fourier Transform (STFT). The K-means algorithm enables precise localization and sizing of surface defects without prior labels, while the gamma process captures the stochastic progression of damage over time. The method demonstrates high accuracy in estimating defect geometry and prognostic trajectories while maintaining low computational complexity. This unified, interpretable approach is generalizable to a range of materials and holds strong potential for real-time structural health monitoring (SHM) in safety-critical domains.