Abstract: During the operation of tunnels, structural defects such as voids, cavities, and loosening are prone to occur in the lining, posing serious threats to operational safety. There is an urgent need for accurate and efficient non-destructive testing technologies. This paper proposes a method integrating Ground Penetrating Radar (GPR) and deep learning to detect rebars and three typical types of defects within tunnel linings. Firstly, simulation models were constructed using GPRMax software based on the Finite-Difference Time-Domain (FDTD) method to simulate the electromagnetic responses of rebars, voids, cavities, and loosening areas, generating a dataset of 2,300 GPR simulation images. Subsequently, a signal and image enhancement process was applied by integrating algorithms including DC removal, wavelet denoising, time-zero correction, segmented gain, and grayscale histogram equalization. Then, by comparing the YOLO series of algorithms, the YOLOv7 network was optimized to build a detection model, and a high-precision base model was obtained through training on the simulation data. Finally, transfer learning was conducted by incorporating actual GPR data collected from a subway tunnel under construction in Beijing to validate the model's practicality. Field detection results demonstrate that the proposed method achieves a precision of 96.5%, a recall of 92.9%, and a mean average precision (mAP) of 94.7% in detecting internal lining defects. This method can effectively achieve intelligent multi-target defect identification, providing reliable technical support for the structural health monitoring of tunnels and showing broad prospects for engineering application.