Abstract: Reliable recognition of internal voids in deep shaft linings is essential for maintaining structural integrity.This study developed an automated recognition approach that integrates rapid forward modeling with an improved YOLOv8 model to recognize void geometries in Ground Penetrating Radar (GPR) images.A two-dimensional shaft model was constructed using gprMax,and void defects with different shapes and burial depths were simulated through the Finite-Difference Time-Domain (FDTD) method.Wavefield and waveform analyses were performed to reveal the electromagnetic response characteristics of lining voids and to establish recognition criteria.A rapid forward modeling algorithm was employed to automatically partition repetitive computational units,enabling the efficient generation of 1 472 simulated GPR images for dataset construction.The improved YOLOv8 model embedded with the CBAM attention mechanism increased the mean Average Precision (mAP) by 93.6%.The results show that when the rebar-void spacing exceeds 0.75m,the secondary recognition confidence rises from 0.29 to a maximum of 0.88,and the confidence for detecting rectangular voids in field GPR data reaches 0.83.These findings demonstrate that combining rapid forward modeling with deep learning provides a high-accuracy and cost-effective solution for non-destructive evaluation of shaft linings,offering a transferable framework for detecting similar underground structural defects.