CNN-based deep learning framework for damage detection and residual load prediction of glass fiber composite laminate

Abstract

Recent advances in convolutional neural networks (CNNs) have significantly enhanced non-destructive investigation (NDI) techniques by enabling automated feature extraction and damage assessment from inspection data. Despite these advances, key challenges remain, including limited availability of labeled datasets, uncertainty in model predictions, and the integration of heterogeneous experimental data for reliable damage quantification and residual performance prediction. In this study, a deep learning framework is proposed for automated damage detection and residual load prediction using infrared thermographic data. A multi-task CNN was trained on a dataset comprising 7688 infrared thermal images acquired during static pure shear tests, representing four specimen conditions: pristine, drilled, 30 mm patch-repaired, and 50 mm patch-repaired. The dataset further incorporates synchronized load and displacement measurements, enabling simultaneous learning of damage- and load-related features. The proposed CNN predicts damage percentage, damaged area, residual load, and specimen category within a unified architecture. The optimized model achieved high predictive accuracy, with an R² value of 0.997 and low RMSE and MAE values, while demonstrating strong generalization across training, validation, and test datasets. Comparative evaluation with pretrained architectures, including ResNet50, EfficientNetB0, and DenseNet121, was conducted. Additionally, uncertainty quantification revealed high confidence in the predictions. The results indicate that CNN-based analysis of thermographic data is a promising approach for non-destructive investigation and residual performance assessment.