论文详情
针对传统地震噪声压制方法存在的泛化性不足、主观性强以及实际无噪声数据稀缺等问题, 利用深度学习方法的泛化特性, 在保护有效信号的基础上, 建立了一种地震噪声智能压制方法。基于有效利用少量实际无噪声数据的原则, 首先通过正演数值模拟地震数据构建数据集, 再搭建基于InceptionV4卷积模块和注意力机制的卷积自编码器网络, 并利用正演数据对网络预训练。该过程首先依靠数据驱动方法和网络强大的特征提取能力初步获取地震数据的有效特征表达, 通过正演数值模拟数据试验分析发现, 该方法既能有效压制绝大部分随机噪声和相干噪声, 还能比DnCNN网络更大程度地避免损伤有效信号; 然后, 再采用迁移学习的策略和少量实际地震数据继续训练网络, 最终获得实际地震数据噪声压制模型。将该方法应用于实际地震数据噪声压制试验, 并从压制效果、保幅性等方面评价方法效果, 结果表明该方法对于随机噪声、面波等噪声干扰具有一定的压制能力, 准确地恢复了有效信号, 且具有处理成本低、效率高等优势。
To solve the problems of insufficient generalization, lack of objectivity, and scarcity of noise-free data in reality in routine denoising methods, we establish an intelligent approach for noise reduction and signal preservation by using the generalization behavior of deep learning. According to the principle of utilizing some observed noise-free data, the data set of synthetic seismogram is first derived from forward modeling, followed by the construction of a convolutional auto-encoding network based on InceptionV4 convolutional module and attention mechanism. The network with great power of feature extraction is pre-trained using synthetic data to tentatively obtain data-driven effective characteristics of seismic data. The tests with modelled data show that our approach attenuates most random noises and coherent noises and is superior to DnCNN in signal preservation. The network is further trained using transfer learning strategy and some observed data to obtain the ultimate denoising model. According to field data tests and performance evaluation from the perspectives of noise reduction and amplitude preservation, our approach is capable of suppressing random noises and surface waves to accurately recover effective signals; it also has advantages in low cost of processing and high efficiency.