In Joo ¹ , Kyeongtae Son ² , Yong-Seong Kim ² , Ga-Ae Ryu ³ and Kwan-Hee Yoo ¹ , ¹Chungbuk National University, Republic of Korea, ²LAT Co.,Ltd, Republic of Korea, ³Korea Institute of Ceramic Engineering & Technology, Republic of Korea

Membrane Electrode Assembly (MEA) quality strongly affects performance and yield in proton exchange membrane fuel cell (PEMFC) manufacturing. This paper presents an end-to-end, production-oriented inspection system that automates MEA feeding, dual-side high-resolution line-scan imaging, and AI-based surface defect decision making. The system accepts 3-/5-layer MEA sheets stacked with interleaving papers in a loader tray, performs separator removal and vacuum fixation, then scans the MEA with coaxial and side lighting to acquire uniform, distortion-free high resolution 8192×11000 images. A hybrid AI model integrates EfficientNetV2-based classification with DETR-based object detection; a quadrant tiling strategy and bilateral-filter+CLAHE preprocessing enable stable inference on the high-resolution images under GPU memory constraints. Evaluation of our system achieved 97.11% accuracy and 39.2 s/part tact time for dual-side inspection.

MEA, line-scan imaging, automated handling, vacuum pick-up, deep learning, EfficientNetV2, DETR, anomaly detection