分布式光纤声传感器（DAS），与传统地震传感器相比，因其组网能力强、作业高效低成本、宽频、抗电磁干扰、耐高温高压等恶劣环境的一系列优势，具有广阔的发展前景，已经成为世界领域地震勘探数据采集新亮点。但是，DAS技术广泛应用仍面临一些挑战，其中之一为接收的数据噪声水平较高，严重降低了采集资料信噪比，给数据处理解释带来较大困难。目前各类去噪算法很多,一个主要瓶颈是都需要一定的前提假设以及适用条件,对DAS数据消噪处理都收效甚微。本项目立意开展基于深度学习的DAS数据噪声压制研究，利用深度学习不依赖于先验假设而是从采集大量样本数据出发的优势：分析DAS数据特性，建立DAS数据噪声源函数及信号与噪声模型；构建适合DAS的深度学习训练集；采用高斯过程的贝叶斯优化方法，获得网络超参数的最优组合；将深度学习与时空域分析相结合，对深度学习网络进行改进，最终建立一种适用DAS地震资料噪声智能压制的方法与技术。

Compared with the traditional seismic sensor, distributed optical fiber acoustic sensor (DAS) has a wide development prospect because of its strong networking ability, high efficiency, low cost, wide band, anti-electromagnetic interference, high temperature and high pressure resistance and so on. It has become a new highlight of seismic exploration data acquisition in the world. However, the wide application of DAS technology still faces some challenges, especially the high level of received noise which seriously reduces the signal-to-noise ratio of the collected data and brings great difficulties to data processing and interpretation. Most existing denoising methods rely on prior hypothesis, so they have little effect in DAS data denoising. This project intends to carry out the intelligent denoising study of DAS data based on deep learning, make full use of the deep learning, which doesn’t rely on prior hypothesis but starts from collecting a large number of sample data: by analyzing the characteristics of DAS data, the noise source function of DAS data and the model of signal and noise are established; construct the deep learning training set suitable for DAS data; the optimal combination of super-parameters is obtained by using Bayesian optimization method of Gaussian process; the deep learning network is improved by the combination of deep learning and spatial-temporal analysis. Finally, an intelligent noise suppression method and technology for DAS seismic data is established.