Deep Learning-Based Terahertz Inspection Technique for Internal Defect Detection in Ceramic, Polymer and Metal Composites

<b>Abstract</b><br/><b>:</b> In this study, internal defects in composites composed of ceramics, polymers, and metals were inspected using terahertz technology in combination with a convolutional neural network (CNN) model and anomaly detection employing a generative adversarial neural network (AnoGAN) model. The defective specimens were intentionally fabricated in the hole and edge regions of the polymer layer, which are vulnerable regions in the composites. The defect region in the polymer layer of the composite was detected in the THz scanning image due to the scattering of THz waves.<br/>The THz scanning image dataset was constructed based on the presence or absence of defects in different regions using a terahertz time-domain spectroscopy (THz-TDS) system. The CNN model was trained to determine the scanning regions (hole, edge) of the composites, while the AnoGAN model was trained to detect defects in the THz images. The CNN model demonstrated a 95% accuracy in classifying the THz scanning regions, and the AnoGAN model achieved a 95% accuracy in detecting defects in the composites. This technique enables automated, real-time inspection for detecting internal defects in semiconductor composite products.<br/><br/><b>Acknowledgment</b><br/>This work was supported by Korea Institute of Energy Technology Evaluation and Planning(KETEP) grant funded by the Korea government(MOTIE)(20212020800090, Development and Demonstration of Energy-Efficiency Enhanced Technology for Temperature-Controlled Transportation and Logistics Center) and this work was supported by This research was also supported by a National Research Foundation of Korea (NRF) grant funded by the Korean Government (MEST) (2021M2E6A1084690)