利用多层感知机的地震数据去噪

Seismic data denoising based on multi-layer perceptron

地震勘探广泛应用于地质构造分析、油气及其他矿产资源勘查等领域。受环境、仪器等因素影响,地震数据中不可避免地混杂随机噪声,无疑会对后续的资料处理和解释带来负面影响。文中提出一种用多层感知机(Multi-layer Perceptron,MLP)的去噪方法:首先用滑动窗口在已知地震数据中采样并将其转换为一维向量,作为训练集样本构建多层神经网络模型;再通过反向传播算法计算模型各层神经元的权重,使得模型训练均方误差最小;然后将合成或实测含噪地震数据输入到已训练模型,用该已训练得到的权重计算模型输出。将上述MLP方法处理结果与曲波方法去噪结果进行比较,发现MLP方法去噪结果的信噪比更高,且较好地保护了有效信号,尤其是对构造细节有显著的保护效果。

Seismic exploration has played an important role in tectonic analysis and prospecting of hydrocarbon and other mineral resources.Due to the influence of environment and instruments,seismic data contain random noises,which have a negative impact on processing and interpretation.We propose a multi-layer perceptron(MLP)method to reduce random noises.Seismic data are sampled using a moving window and then converted into a 1 D vector,which is utilized as training samples to establish a multi-layer neural network model.The weighting factor of neurons in each layer is calculated using the back propagation algorithm until the mean square training error reaches a minimum.Synthetic or measured noisy seismic data are imported into this established model,and the output is calculated using the weighting factors after training.We compare the denoising results derived from MLP and curvelet methods and conclude that MLP result exhibits higher signal to noise ratio and better signal preservation,especially for structural details.