| Abstract [eng] |
Composite materials are widely used in lightweight and safety-critical structures; however, due to their layered architecture, they are sensitive to various types of damage. Defects such as notches or interlaminar delamination can alter the strain field and reduce structural reliability. Therefore, it is important not only to detect damage, but also to quantitatively evaluate its parameters. The aim of this work is to develop and investigate a machine learning methodology for detecting, classifying and evaluating defects in a composite plate using finite element model-generated synthetic data and experimental Digital Image Correlation strain fields. Two defect types are analysed: a notch and delamination. Five tasks are formulated: notch detection, delamination detection, notch depth prediction, notch angle prediction and delamination length prediction. Since the amount of experimental data with known defect parameters is limited, synthetic data generated using a parametrized finite element model were used for model training. Defect parameter combinations were defined using a Design of Experiments approach. The generated strain fields were interpolated into regular grids, segmented and transformed into statistical feature vectors, which were then used for classical supervised machine learning models. The results showed that the larger dataset of 328 numerical cases significantly improved regression accuracy compared with the 113-case dataset. Based on the best independent validation results, mean absolute error (MAE) decreased by 48.3% for notch depth prediction, 54.7% for notch angle prediction and 46.0% for delamination length prediction. It was also found that segmented features were more informative than global features in regression tasks, with the best case showing an 82.6% reduction in MAE. The analysis of models and hyperparameters showed that different regression tasks require models of different complexity, and that accuracy improvements must be evaluated together with computational cost. The developed methodology was tested using experimental DIC strain fields. The models correctly identified the notch, while its angle and depth were predicted with accuracies of 94.7% and 89.8%, respectively. |
