Microseismic Denoising and Reconstruction by Unsupervised Machine Learning

Microseismic data reconstruction is a procedure to compensate for acquisition deficiencies and to improve the data quality, which is important for subsequent processing steps such as event location. The performance of most reconstruction methods depends on 1) their parameter settings and 2) degrades greatly in case of strong noise interference. We propose an unsupervised machine learning algorithm to realize the incomplete noisy data reconstruction, using the Indian Buffet Process (IBP) as a prior to learning an appropriate dictionary from the noisy data. An approximation to the full posterior is obtained via Gibbs sampling, yielding an ensemble of dictionary and sparse coefficients. Finally, the signal of interest is reconstructed by the product of the dictionary and sparse coefficients. Tests on synthetic and real microseismic data demonstrate that the proposed method works very well for low signal-to-noise ratio data with missing traces.