Seismic noise attenuation by signal reconstruction: an unsupervised machine learning approach

Random noise attenuation is an essential step in seismic data processing for improving seismic data quality and signal-to-noise ratio. We adopt an unsupervised machine learning approach to attenuate random noise via signal reconstruction strategy. This approach can be accomplished in the following steps: Firstly, we randomly mute a part of the input data of the neural network according to a certain percentage, and then the network outputs the reconstructed data influenced by this randomly mute. The objective function measures the distance between the input data and the reconstructed data. Secondly, we use the adaptive moment estimation algorithm to minimize the distance, and the network adjusts its internal parameters so that sparse representations can be captured by the multiple processing layers of the neural network. Finally, we take the same proportion of random mute on the raw seismic data which are fed to the trained neural network. Through this network, reconstruction of seismic data and attenuation of random noise are completed simultaneously. We use both synthetic and field data to testify the feasibility and applicability of the proposed method. Synthetic data experiment indicates that the proposed method achieves better denoised results than the conventional methods. Field data applications further demonstrate its superiority and practicality.