Noise reduction for micro-seismic monitoring using deep learning
AssigneeExxonMobil Technology and Engineering Company
InventorsSudeep Singh, Xinlong Liu, Simona O. Costin, Christopher T. Reaume, Taylor K. Fink
Life of this patent
Estimating the term…
Abstract
A method for classifying a microseismic event, including: analyzing microseismic event files through a combination of two fault tolerant machine learning pipelines, an acoustic machine learning pipeline and a visual machine learning pipeline; and generating a classification prediction for the microseismic event files by combining predictions from the acoustic machine learning pipeline and the visual machine learning pipeline.
Claims10
A method for classifying a microseismic event, comprising:
- analyzing, in a real-time manner and via one or more multi-layer neural networks, microseismic event files through a combination of two fault tolerant machine learning pipelines, wherein the two fault tolerant machine learning pipelines include:
- at least one of an acoustic machine learning pipeline, the acoustic machine learning pipeline including:
- dividing the microseismic event files into a plurality of component audio files;
- converting the plurality of component audio files into power spectrum files;
- and
- generating a plurality of predictions for the power spectrum files by applying an acoustic machine learning model corresponding to each component used to divide the microseismic event files into the plurality of component audio files;
- and
- at least one of a visual machine learning pipeline, the visual machine learning pipeline including:
- converting the microseismic event files into plurality of spectrograms;
- generating a prediction for a fused spectrogram representation by applying a visual machine learning model;
- and
- converting the spectrograms into RGB images, and then generating an additional prediction for the spectrograms by applying the visual machine learning model to the RGB images;
- and
- generating a classification prediction for the microseismic event files by combining predictions from the acoustic machine learning pipeline and the visual machine learning pipeline.
A non-transitory computer readable storage medium, encoded with instructions, which when executed by a processor causes the processor to implement a method for classifying a microseismic event, the method comprising:
- analyzing, in a real-time manner and via one or more multi-layer neural networks, microseismic event files through a combination of two fault tolerant machine learning pipelines, wherein the two fault tolerant machine learning pipelines include:
- at least one of an acoustic machine learning pipeline, the acoustic machine learning pipeline including:
- dividing the microseismic event files into a plurality of component audio files;
- converting the plurality of component audio files into power spectrum files;
- and
- generating a plurality of predictions for the power spectrum files by applying an acoustic machine learning model corresponding to each component used to divide the microseismic event files into the plurality of component audio files;
- and
- at least one of a visual machine learning pipeline, the visual machine learning pipeline including:
- converting the microseismic event files into plurality of spectrograms;
- generating a prediction for a fused spectrogram representation by applying a visual machine learning model;
- and
- converting the spectrograms into RGB images, and then generating an additional prediction for the spectrograms by applying the visual machine learning model to the RGB images;
- and
- generating a classification prediction for the microseismic event files by combining predictions from the acoustic machine learning pipeline and the visual machine learning pipeline.
Dependencies are inferred from the claim text.