Removal of powerline noise in geophysical data sets with a scientific machine-learning based approach

Research output: Contribution to journal/Conference contribution in journal/Contribution to newspaperJournal articleResearchpeer-review

Abstract

The most common noise in geophysical data is probably interference from powerlines. This noise manifests itself as a sinusoidal signal oscillating at the fundamental 50 or 60 Hz frequency of the power grid and as harmonic components oscillating at integer multiples. Many different mitigation strategies, tailored for the specific geophysical method, have been developed to target powerline noise. One method that applies to fully sampled data is model-based subtraction, where a model of the powerline noise is fit to the noisy dataset and subsequently subtracted. In most cases, this leads to significant improvements in the signal-to-noise ratio. However, the determination of the powerline model parameters, in particular the fundamental powerline frequency, is computationally expensive, as it requires repeated solutions of a least-squares problem. We demonstrate that the powerline frequency can be directly predicted with a scientific machine-learning-based approach. We work on both time domain-induced polarization and surface nuclear magnetic resonance data. We use a different network for each method to trade-off prediction accuracy and prediction speed. In both cases, the prediction accuracy is fully on par with standard methods, and we obtain speed-ups by factors of 400 and 10 for the two types of data.

Original languageEnglish
Article number5923410
JournalIEEE Transactions on Geoscience and Remote Sensing
Volume60
Number of pages10
ISSN0196-2892
DOIs
Publication statusPublished - 2022

Keywords

  • Frequency prediction
  • geophysical datasets
  • powerline noise
  • scientific machine learning
  • surface nuclear magnetic resonance (NMR)
  • time domain-induced polarization (TDIP)

Fingerprint

Dive into the research topics of 'Removal of powerline noise in geophysical data sets with a scientific machine-learning based approach'. Together they form a unique fingerprint.

Cite this