基于变分自动编码器的探地雷达噪声压制及双曲线识别

Ground penetrating radar noise suppression and hyperbolic recognition based on variational autoencoder

受仪器设备和环境的影响,使用探地雷达(GPR)进行数据采集的过程中会引入各种不同的噪声,从而干扰目标体反射信号,对数据处理及解释造成困难.由此,本文提出了一种能够同时实现GPR剖面数据噪声压制及双曲线异常识别的变分自动编码器(VAE)神经网络结构.首先,详细阐述了模型结构、数据集建立及网络参数选择;然后,通过合成数据实验验证该算法的有效性及模型的泛化能力,相比于深度卷积自动编码器(CDAE),本文算法在噪声压制能力和双曲线识别能力上均表现更优;最后,通过对隧道衬砌和管线的实测数据进行处理,验证了本文算法的实用性.

Due to the influence of instrumentation and environment, various kinds of noise are introduced in the process of data acquisition using Ground Penetrating Radar(GPR), which interferes with the reflected signals of the target body and causes difficulties in data processing and interpretation. In this paper, we propose a Variational Auto Encoder(VAE) neural network structure that can simultaneously achieve noise suppression and hyperbolic anomaly recognition in GPR profile data. First, the model structure, data set establishment and network parameter selection are described in detail. Then, the effectiveness of the proposed algorithm and the generalization ability of the model are verified by the synthetic data experiments. Compared with the Convolutional Denoising Autoencoder(CDAE), the proposed algorithm has better noise suppression ability and hyperbolic recognition ability. Finally, the practicability of the proposed algorithm is verified by processing the measured data of tunnel lining and pipeline.