ScholarWorks@동국대학교

초록

Structural Health Monitoring (SHM) of laminated composite materials using Lamb Waves (LW) provides an efficient means for early damage detection and structural integrity assessment. However, traditional data-driven methods depend on manual feature extraction, and standard Convolutional Neural Networks (CNNs) struggle to capture the complex temporal–spectral characteristics of LW signals. To overcome these limitations, this study introduces a Parallel Kernel Convolutional Neural Network (PK–CNN) framework for automated and precise damage detection along with localization in laminated composite materials. In this study, the tasks were trained using two PK–CNN classifiers that share the unified PK–CNN backbone, having the same parallel-kernel feature extractor, with separate classification heads trained for each task. The severity model outputs four classes (H, D1, D2 and D3) while the localization model outputs nine classes (L1–L9). Both models take the same LW input signals. The PK–CNN employs parallel convolutional layers with varying kernel sizes to extract both locally and globally scaled features simultaneously from LW signals. Systematic hyperparameter tuning was performed to ensure optimal convergence and network stability. Experimental validation on LW datasets from laminated composite plates demonstrated that the PK–CNN achieved 95.83% test accuracy for damage severity classification and 98.15% accuracy for localization across nine zones. Quantitative metrics confirmed high precision, recall, and F1-scores of (0.96, 0.96, 0.96) for severity classification and (0.98, 0.98, 0.98) for localization. The PK–CNN eliminates dependency on kernel-specific tuning, effectively learns narrow- and broad-scale features, and enables a unified architecture for simultaneous damage severity assessment and spatial localization in composite SHM applications. © 2026 Elsevier Ltd.

키워드