A real-time and offline weld defect detection framework using planar EMT with selective measurement and deep learning

Electromagnetic tomography (EMT) offers a non-contact approach for weld defect imaging, yet traditional full-pair configurations suffer from low speed and limited adaptability to uneven weld surfaces. This study proposes an integrated real-time and offline weld-defect detection framework that couples selective coil-pair planar EMT with a prototype-guided similarity-fusion UNet (ProSiF-UNet). A 16-channel planar EMT sensor was developed, and a simplified sensitivity matrix using next-nearest coil pairs (d = 2) significantly reduced computation while preserving image fidelity. For real-time monitoring, a lightweight linear back projection (LBP)-based reconstruction enables 200+ frames per second (FPS) imaging, whereas an offline Landweber (LW) iteration ensures high-fidelity analysis. The prototype-guided similarity-fusion segmentation network (ProSiF-UNet) integrates cosine-similarity prototypes and exponential moving average (EMA)-based fusion to enhance defect segmentation on limited data, reducing missed detections by 40% compared with standard U-Net. The unified framework achieves a practical balance between speed, accuracy, and robustness, demonstrating strong potential for artificial intelligence (AI)-enabled industrial weld inspection.