Multipatch Deep Learning for Multilevel Damage Assessment of Carbon-Fiber-Reinforced Polymer Plates from Lamb-Wave Continuous Wavelet Transform Images

To ensure the reliability of carbon-fiber-reinforced polymer (CFRPs) structures, robust structural health monitoring (SHM) is required for timely damage diagnosis. Over the decades, Lamb-wave-based techniques have been widely used to inspect composite structures owing to their high sensitivity to internal defects. Damage-sensitive features from nonstationary Lamb-wave signals can be captured in the frequency and time domains using the continuous wavelet transform (CWT). However, extracting localized damage features from these images is highly challenging. Accordingly, this study proposes a multipatch deep feature learning method for damage detection and severity classification in CFRP plates using Lamb-wave CWT images. The images are partitioned into multiple local patches for patch-wise deep feature extraction using a convolutional neural network (CNN). The proposed model is evaluated on CFRP composite plates with three simulated damage severity levels (D1, D2, and D3) produced using mass blocks of different weights. Among the evaluated patch configurations, the proposed physics-inspired patching achieves the highest classification accuracy of 98.2%, outperforming conventional uniform multipatch baselines. For damage detection, the proposed method achieves high classification performance, with precision, recall, and F1-scores of 100% for both healthy and damaged samples, outperforming comparison models, including a custom CNN, VGG19, and ResNet50. For damage severity classification, the proposed model achieves F1-scores of 0.98, 0.97, and 0.98 for D1, D2, and D3, respectively, consistently outperforming the baseline models across various evaluation metrics. Under Gaussian noise, the proposed method maintains a robust classification accuracy of 96.1% at 20 dB signal-to-noise ratio, corresponding to a performance reduction of 1.7% compared with noiseless data, suggesting its reliability for realistic SHM environments.