Ground-penetrating radar (GPR) is a nondestructive near-surface geophysical detection method, which is often used to locate subgrade diseases in railway subgrade detection. However, the strong noise reflected by the sleeper will obscure the effective information of the subgrade disease. In order to suppress this strong noise, the traditional processing method is generally filtering, which is dependent on expert experience. We propose a method of railway sleeper interference suppression based on UNet. We use datasets based on real railway subgrade structures to train UNet. Each dataset consists of a pair of GPR data with sleeper interference and GPR data without sleeper interference. The experimental results of simulation data show that the SSIM value between the data processed by UNet and the data without sleeper interference exceeds 0.99. We test our approach on field data collected on the Qinghai-Tibet Railway. Compared with traditional filtering methods, UNet does not rely on expert experience and is more effective. The research shows that our method can effectively suppress the sleeper strong interference in GPR data and improve the signal-to-noise ratio. We expect that our method can promote the development of railway subgrade detection.

中文翻译：

---

基于深度学习的铁路路基检测探地雷达数据强噪声抑制


探地雷达（GPR）是一种无损近地表地球物理探测方法，常用于铁路路基检测中的路基病害定位。但枕木反射的强烈噪声会掩盖路基病害的有效信息。为了抑制这种强噪声，传统的处理方法一般是滤波，这依赖于专家经验。提出一种基于UNet的铁路枕木干扰抑制方法。我们使用基于真实铁路路基结构的数据集来训练 UNet。每个数据集由一对有轨枕干扰的 GPR 数据和无轨枕干扰的 GPR 数据组成。仿真数据实验结果表明，UNet处理后的数据与无轨枕干扰的数据之间的SSIM值超过0.99。我们在青藏铁路上收集的现场数据测试了我们的方法。与传统的过滤方法相比，UNet不依赖专家经验，更加有效。研究表明，我们的方法可以有效抑制探地雷达数据中的卧铺强干扰，提高信噪比。我们希望我们的方法能够促进铁路路基检测的发展。