2D convolutional neural network and laser ultrasonic imaging for delamination detection in CFRP laminates

Abstract Carbon Fiber-reinforced polymer (CFRP) composites are widely used in safety-critical applications, where early detection of damage like delamination is essential. Ultrasonic Guided Waves (UGWs) are highly effective for assessing damage due to their ability to propagate over long distances with minimal attenuation and high sensitivity to defects. However, their multimodal and dispersive behaviour complicates interpretation, particularly in anisotropic materials such as CFRPs. This study explores the use of deep learning (DL) to automate delamination detection in quasi-isotropic CFRP laminates. Ten CFRP plates were fabricated, one undamaged and nine with artificial delamination of varying sizes and depths. UGW data were collected using a laser ultrasonic setup across three sensor positions. A 2D Convolutional Neural Network (2D-CNN) was developed to classify delamination from wavefield images. The model’s performance was tested across three training-validation-test splits: 70%-20%-10%, 50%-20%-30%, and 30%-20%-50%. The 2D-CNN achieved test accuracies of 93.5%, 89% and 78.4% respectively. Confusion matrices confirmed low misclassification rates, especially in larger training sets. Although performance declined with reduced training data, the model consistently showed strong generalization, demonstrating the potential of DL for reliable, automated non-destructive evaluation in composite structures.