Low-frequency signal and noise are the main components of seismic data in Northwest China desert area. To remove noise and preserve data details, we introduce a method of nonlinear autoregressive with external input neural network (NARX) which is based on recent developments in nonlinear time series prediction and improve its performance according to signal detection theory. The input of NARX is a noisy seismic signal, the output is original predicted background noise, and then a threshold is set for the output to eliminate the waveform distortions. The residual error of the input and the output is filtered signal. We test the proposed method on synthetic and field seismic data and compare it with some conventional filtering methods (wavelet denoising method, f–x filter and complex diffusion filtering). The results prove that the proposed method can greatly attenuate low-frequency noise, and preserve data details as completely as possible.

低频信号和噪声是西北荒漠地区地震资料的主要成分。为了去除噪声并保留数据细节，我们引入了一种基于非线性时间序列预测的最新发展的带有外部输入神经网络（NARX）的非线性自回归方法，并根据信号检测理论提高了其性能。NARX的输入是有噪声的地震信号，输出是原始预测的背景噪声，然后为输出设置一个阈值以消除波形失真。输入和输出的残差为滤波信号。我们在合成和现场地震数据上测试了所提出的方法，并将其与一些传统的滤波方法（小波去噪方法，f - x滤波和复扩散滤波）。结果证明，该方法可以极大地衰减低频噪声，并尽可能完整地保留数据细节。