基于迁移学习的隧道凿岩台车钻孔震源实测地震数据去噪方法及应用

Noise Reduction Method and Application of Drilling Source Seismic Data for Tunnel Rock Drilling Rig Based on Transfer Learning

隧道凿岩台车钻孔震源实测信号具有数据杂乱、噪声干扰严重及偶发性采集异常等特点,难以直接用于模拟数据训练后的神经网络模型.针对这一问题,首先基于U-Net去噪神经网络,将互相关、Automatic Gain Control(AGC)和归一化处理作为U-Net神经网络的预处理层,开展信号降噪处理,有效提升了凿岩台车含噪钻孔震源与无噪地震记录间复杂映射关系的拟合效果.然后,构建了基于Generative Adversarial Nets(GAN)的对抗式迁移学习神经网络,将隧道的真实地震数据进行迁移,实现了真实数据在已训练完成的U-Net去噪神经网络中的有效应用.最后,依托实际工程对该方法进行评价.结果表明,采用提出的去噪方法,可有效去除地震记录中的噪声干扰,探测结果与实际开挖情况基本吻合,初步验证了该方法的可行性与有效性.

Aiming at the problem that the measured signals from the borehole seismic source of tunnel rock drilling rigs are characterized by data clutter,serious noise interference and occasional collection anomalies,which make it difficult to be directly applied to the neural network model trained with the simulated data.In this paper,based on the U-Net denoising neural network,mutual correlation,Automatic Gain Control(AGC)and normalization are used as the preprocessing layer of the U-Net neural network to carry out the signal noise reduction process,which effectively improves the fitting effect of the complex mapping relationship between the noisy borehole source and the noiseless seismic recordings of rock drilling cart.Then,an adversarial transfer learning neural network based on Generative Adversarial Nets(GAN)was constructed to migrate the real seismic data from the tunnel,realizing the effective application of real data in the trained U-Net denoising neural network.Finally,relying on the actual project to evaluate the method,the results show that,using the denoising method proposed in this paper,the noise interference in the seismic record can be effectively removed,and the detection results are basically in line with the actual excavation,which preliminarily verifies the feasibility and effectiveness of the method.