Enhancing Crack Detection in Critical Structures Using Machine Learning and 3D Digital Image Correlation

Three-dimensional digital image correlation (3D-DIC) is a non-contact monitoring technique that is able to provide accurate three-dimensional strain and displacement measurements. Previous research has shown that 3D-DIC can detect micron-scale cracks in structures as they emerge; however, because 3D-DIC is an optical sensing technique, unfavorable visual conditions due to high heat, large deformations, or a significant distance between the structure and the 3D-DIC cameras can make crack detection difficult or impossible.

This research aims to develop machine learning algorithms capable of detecting characteristic crack signals in these scenarios.

Localized point velocities obtained via 3D-DIC were transformed into 2D color images for machine learning segmentation. A novel dataset processing technique was utilized to produce the training dataset, which overlayed simplistic crack analogs on top of the first 50 images from the test. Different parameters from this technique were investigated to determine their effect on the model’s accuracy and sensitivity.

The resulting model detected the onset of significant cracking with an accuracy comparable to acoustic emissions sensors. Varying the processing parameters yielded models that could detect evidence of cracking earlier, at the cost of potentially higher false positive rates. The model also performed well on structures imaged in similar testing setups that were not included in the training dataset.

This data processing technique enables crack detection in scenarios where acoustic emissions and other sensors cannot be used. It additionally allows processes already utilizing 3D-DIC to obtain additional information about material performance during testing or operation.