1D convolutional neural network for delamination detection in CFRP laminates using laser ultrasonic guided waves

Among conventional non-destructive evaluation methods, Ultrasonic Guided Waves (UGWs) have proven highly effective for evaluating damage in fiber-reinforced polymers (FRPs). This is attributed primarily to their capacity to propagate over extended distances with minimal attenuation, coupled with their high sensitivity to material discontinuities. However, the multimodal nature and dispersion of UGWs complicate signal interpretation, particularly in anisotropic materials where wave interactions become more intricate. Consequently, the utilization of UGWs for structural health monitoring (SHM) frequently necessitates extensive post-processing to accurately interpret the data and differentiate between overlapping wave modes. The process is often laborious, costly, and vulnerable to bias. In order to address this challenge, researchers are increasingly turning to deep learning (DL) to automate feature extraction from complex datasets. Convolutional neural networks (CNNs), a specific subset of DL models, excel at autonomously learning hierarchical spatial features. This renders them particularly effective for the detection of subtle damage signatures in wavefield patterns. The necessity for reliable diagnostic tools to assess high-performance FRPs under stringent safety standards is the primary motivation for this study. Its aim is to develop CNN-based methods to detect and characterize delamination in carbon fiber-reinforced polymer (CFRP) laminates using raw laser-induced Lamb waves. An extensive experimental campaign was conducted using a laser ultrasonic visualization inspector (LUVI-CP1) and a broadband piezoelectric transducer to collect UGW signals under various delamination scenarios. Cross-ply [902/02]S and quasi-isotropic [0/45/90/-45]S CFRP laminates were manufactured with artificially induced delamination (5x5 mm2, 10x10 mm2 and 20x20 mm2), each at three through-thickness positions and recorded from three sensor positions. The dataset included both damaged and undamaged laminates, totaling over 3,750,000 signals for each laminate type. A lightweight one-dimensional CNN (1D-CNN) was developed to process each UGW trace as a time-series vector. Compared with 2-D CNNs, the 1-D approach preserves classification accuracy while significantly reducing the model parameters, inference latency, and training-data requirements. The architecture includes a single convolutional block, two fully connected layers, and a terminal SoftMax layer that classifies each input by its unique combination of laminate lay-up, delamination size, through-thickness position, and wave propagation direction. Model generalizability and performance were quantified through stratified k-fold cross-validation. First, the network was trained on k-1 folds and evaluated on the hold-out fold, cycling until every subset had served once as the test set. A complementary protocol then inverted the split, training on a single fold and testing on the remaining k-1 folds to ensure robustness under severe data scarcity. Across both protocols, the 1D-CNN delivered 99.9 % accuracy in validation and test phases, confirming its reliability for automated delamination classification and highlighting the promise of compact DL models for practical SHM.