PSI - Issue 68

Giovanni Chianese et al. / Procedia Structural Integrity 68 (2025) 1245–1251 Chianese et al. / Structural Integrity Procedia 00 (2025) 000–000

1250

6

3.2. Approach 2 strain fields measurements processed with CNNs A subset of strain fields along the Y direction, ε Y , collected during simulations and referred to the region of interest, is reported in Fig. 4. Their analysis revealed that at each value of the normalized crack length α , different crack front geometries resulted in different contour plots of the strain fields ε Y . Moreover, when the crack geometry varied from a straight front to an elliptical one (increasing β ), the bean-shaped region interested by higher values of the strain increased its size and bent over itself. For this reason, the spatial distribution of the strain on the lateral surface of the specimen indicated that full field measurements seem to provide useful information for classification of different crack front aspect ratios. Analogously, it was assumed that additional information would be carried by the data representing the strain field along the X direction, ε X , which, therefore, was stored in the dataset along with ε Y . This suggested a strategy for implementation of an automatic system since convolutional neural networks enable automatic processing of images. For this reason, strain fields ε X and ε Y were concatenated and treated as images before being processed by two CNNs to simultaneously predict α and β . Similar approaches were investigated by Melching et al. (2022) to detect the crack tip in fatigue crack growth, by Pal et al. (2023) for the localization of subsurface damage, and by Li et al. (2022) for structural health monitoring systems.

Fig. 4. Contour plot of ε Y (m/m) in the ROI of the specimen (the red square at the top-right of the image).

Fig. 5. (a) Normalized crack length prediction with CNN; (b) confusion chart with classification of the crack tunnelling conditions with a CNN.