Inflection point principle combined with digital image correlation and machine learning for crack length measurement in fatigue tests

• Published in: Theoretical and applied fracture mechanics Vol. 139; p. 105052
• Main Authors: Ščerba, Bořek, Adamec, Tomáš, Pokorný, Pavel, Návrat, Tomáš, Vajdák, Michal, Náhlík, Luboš
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 01.10.2025
• Subjects: Crack length measurement; Digital image correlation; Gaussian process regression; Inflection point method; Machine learning; Crack length measurement; Gaussian process regression; Inflection point method; Digital image correlation; Machine learning
• ISSN: 0167-8442
• DOI: 10.1016/j.tafmec.2025.105052

•A novel Inflection Point Method (IPM) for real-time fatigue crack length measurement.•IPM integrates contactless digital image correlation (DIC) and machine learning.•Eliminates threshold dependence typical for other DIC-based techniques.•Validated via visual inspection applied uniquely on the same side of the sample.•Outperforms standard DIC line-based thresholding technique. The visual inspection method is a widely used non-contact technique for measuring fatigue crack propagation, but it is inefficient, requiring frequent operator input. Digital image correlation (DIC) methods provide alternatives. However, full-field methods are computationally demanding, while line-based thresholding techniques are sensitive to material load conditions, reducing consistency. This study proposes and validates a new non-contact, physically-based method for real-time crack length evaluation. It eliminates the need for thresholding and enables higher testing frequencies due to its line-based nature, supporting accurate, versatile, and automated fatigue testing. The method integrates the inflection point principle with DIC and machine learning. Visual inspection serves as a validation baseline, using a novel setup that applies both methods simultaneously on the same side of the sample for direct comparison. The proposed method shows good agreement with baseline results, achieving mean absolute errors of 24 μm (static) and 54 μm (dynamic). Compared to line-based thresholding, it is four times more accurate (dynamic) and independent of load levels, though 1.7 times slower.