Investigation of deep learning approaches for automated damage diagnostics in fiber metal laminates using Detectron2 and SAM

The impact damage is one of the major causes of structural failures in Fiber Metal Laminate (FML) plates, which are widely used in the aerospace and automotive industries due to their superior mechanical properties. Accurate detection, segmentation, and characterization of these damages are crucial for improved safety and reduced maintenance costs. This study proposes an automated approach to detect, segment, reconstruct, and characterize the damages in FML plates using state-of-the-art deep learning models: the Segment Anything Model (SAM) and the Mask Region-based Convolutional Neural Network (Mask R-CNN) implemented by the Detectron2 framework. A domain-adapted supervised learning process was applied to the X-ray CT dataset of damaged FML plates impacted with energies of 5J, 7.5J, 10J, and 12.5J. Mask R-CNN significantly outperformed SAM across all key performance metrics while offering around 8 times faster training and 80 times faster inference. Mask R-CNN also proved to have superior explainability for end-users. The lack of absolute ground truth data severely limits the scope of an absolute quantitative comparison, therefore highlighting the need for further studies. This study not only contributes to the area of damage diagnostics in composite materials but also provides insights into the comparative performance and explainability of advanced deep learning models, paving the way for applications in industrial inspection and quality assurance.