复杂不良地质隧道超前GPR图像智能反演方法OA

To address the limitations of traditional empirical interpretation methods for advanced ground penetrating radar(GPR)detection images of tunnels,including inconsistent standards,low efficiency,and high rates of misinterpretation and missed detection,this study proposed an intelligent inversion method for radar B-scan waveform images of adverse geology based on an improved generative adversarial network(GAN).A multi-condition synthetic dataset was constructed through electromagnetic numerical simulation combined with random generation constraints of medium parameters,incorporating three types of geological anomalies(irregular cavities,fracture zones,and cracks)and their combinations.The GAN generator was optimized by integrating Unet convolutional neural network architecture and dilated convolution modules,and structural similarity(SSIM)loss function was incorporated into the loss function to enhance the extraction capability of waveform image features and model inversion performance.Training evaluation results demonstrate that the improved GAN model achieves approximately 4%and 10%higher inversion accuracy than conventional GAN and standalone Unet models,respectively.An inversion experiment using actual engineering data from fracture zone prediction was carried out.Comparative analysis with digital drilling and excavation verification data confirms that the inversion prediction results of the proposed method are basically consistent with actual conditions,and it can provide reliable geological forecasting for tunnel excavation and support.