Ensemble empirical mode decomposition and stacking model for filtering borehole distributed acoustic sensing records

Abstract

We have evaluated the ensemble empirical mode decomposition (EEMD) and stacking model for borehole seismic-data denoising. The borehole records collected by distributed acoustic sensing (DAS) technology have multitype noise contamination, and it is difficult to attenuate these noises while recovering the seismic waves well. We first perform EEMD on the seismic data to obtain the signal-to-noise modal components, then extract the time and frequency information of the decomposed modes using six feature factors, and finally introduce an ensemble learning method to classify the acquired modal features effectively. Stacking is the ensemble learning technique we used in our study. This technique integrates several diverse basic ensemble models using the meta-learning strategy and constructs a highly integrated framework with superior performance and good generalization. In addition, the basic ensemble models consist of many decision tree classifiers following two different ideas of parallelization and serialization. The feature extraction process provides sufficient DAS feature data for the training process of the framework. Synthetic and real experimental results demonstrate that the stacking integration framework effectively separates the signal-to-noise modal features of the borehole DAS records. Furthermore, the EEMD-stacking method performs better than wavelet transform, intrinsic time-scale decomposition, robust principal component analysis, k-means singular value decomposition, and median filtering on the denoising task.