基于时频分析与分数阶最优控制网络的地震数据噪声压制

在复杂的地表环境, 地震勘探采集到的实际地震资料信噪比较低, 分辨率较差, 接收的噪声能量较强, 与有效信号存在频谱的重叠.常规的消噪手段很难在保证有效信号幅值的同时, 还兼顾噪声压制的效果.本文采用基于分数阶最优控制(Fractional Optimal Control)理论建立的深度学习神经网络——FOCNet来压制地震数据噪声, 并恢复微弱同相轴.不同于传统深度学习网络(DCNN)算法大多为基于经验的网络设计, FOCNet具有坚实的数理基础, 它从动态系统的最优控制角度阐述了网络的原理, 并采用长期记忆的方式增强了网络的稳定性, 提高了系统对噪声的消减能力.针对地震数据有效信号在低频带与噪声重叠严重, 且FOCNet对数据中、高频信息保留更好这一情况, 本文提出了一种基于理想时频分析与FOCNet相结合的算法(TF-FOCNet)来压制地震噪声, 提取有效信号.该算法通过理想时频分析, 针对性的提取信噪重叠的低频目标数据成分, 并与数据的中、高频成分一起送入网络中进行处理并融合, 完成噪声的压制, 增强了低频信息的保留能力.模拟及实际的实验结果验证了算法在随机噪声、面波压制及微弱信号恢复上的有效性和优越性.

Fractional optimal control network based on time-frequency analysis for noise suppression of seismic data

In the process of real seismic exploration, the collected seismic data has low signal-to-noise ratio and poor resolution due to the complex geographical environment. The received noise has a partial overlap of the frequency spectrum with the effective signals. It is difficult for conventional noise elimination methods to ensure the effective signal amplitude while also taking into account the effect of noise suppression. In this paper, a deep learning neural network-FOCNet based on fractional optimal control theory is used to suppress seismic data noise and recover the weak seismic events. Unlike traditional DCNN methods that are mostly empirical-based in network design, FOCNet has a solid mathematical foundation. It expounds the network principle from the perspective in optimal control of dynamic systems and uses long-term memory mode to enhance network stability and improve system ability to reduce noise. In view of the fact that the effective signals of seismic data overlaps seriously with noise in the low-frequency band, this paper presents an method based on the combination of ideal time-frequency analysis and FOCNet (TF-FOCNet) to suppress seismic noise. The method specifically extracts the target data components of the low-frequency band with ideal time frequency analysis, and sends it into the network for processing and fusion together with the mid/high-frequency components of data, so as to complete the suppression of noise, and enhance the ability to retain low-frequency information. Synthetic and field experiments verify the effectiveness and superiority of the method in random noise and surface wave suppression and weak signals recovery.