Multi-scale deep network for fault detection

Patent metadata

Field	Value
Publication number	US-11467300-B2
Application number	US-201716321030-A
Country	US
Kind code	B2
Filing date	Aug 3, 2017
Priority date	Aug 3, 2016
Publication date	Oct 11, 2022
Grant date	Oct 11, 2022

Abstract

A method for detecting an unknown fault in a target seismic volume. The method includes generating a number of patches from a training seismic volume that is separate from the target seismic volume, where a patch includes a set of training areas, generating a label for assigning to the patch, where the label represents a subset, of the set of training areas, intersected by an known fault specified by a user in the training seismic volume, training, during a training phase and based at least on the label and the training seismic volume, a machine learning model, and generating, by applying the machine learning model to the target seismic volume during a prediction phase subsequent to the training phase, a result to identify the unknown fault in the target seismic volume.

First claim

What is claimed is:

1. A method for detecting an unknown fault in a target seismic volume, comprising:
generating a plurality of patches from a training seismic volume that is separate from the target seismic volume, wherein a patch of the plurality of patches comprises a set of training areas, and wherein the plurality of patches comprises contiguous portions of multiple pixels of the training seismic volume;
generating a label for assigning to the patch, wherein the label represents a subset, of the set of training areas, intersected by a known fault specified by a user in the training seismic volume;
training, during a training phase and based at least on the label and the training seismic volume, a machine learning model; and
generating, by applying the machine learning model to the target seismic volume during a prediction phase subsequent to the training phase, a result to identify the unknown fault in the target seismic volume.

2. The method of claim 1 , further comprising:
obtaining seismic data corresponding to a subterranean formation, wherein the target seismic volume comprises the seismic data;
analyzing the seismic data using the machine learning model to generate the result; and
determining, based on the result, a location of the unknown fault in the subterranean formation.

3. The method of claim 1 , further comprising:
extracting one or more training seismic sections from the training seismic volume wherein the plurality of patches is generated from the one or more training seismic sections,
wherein the known fault is specified in the training seismic volume by the user.

4. The method of claim 1 , wherein generating the label comprises:
identifying a pre-determined label of each training area in the set of training areas; and
including the pre-determined label of each training area in the subset into the label of the patch,
wherein said each training area in the subset comprises a plurality of pixels of the training seismic volume that are intersected by the known fault.

5. The method of claim 1 , further comprising:
obtaining seismic data corresponding to a subterranean formation, wherein the target seismic volume comprises the seismic data;
analyzing the seismic data using the machine learning model to generate the result; and
determining, based on the result, a location of the unknown fault in the subterranean formation,
wherein applying the machine learning model to the target seismic volume comprises:
placing, for an iteration of a plurality of iterations, a sliding window over a seismic section of the seismic data, wherein the machine learning model is applied, during the iteration and
using the sliding window, to the seismic section to generate a partial result, and

wherein generating the result comprises:
aggregating a plurality of partial results generated during the plurality of iterations to generate a heat map,
wherein the result comprises the heat map.

6. The method of claim 1 , further comprising:
obtaining seismic data corresponding to a subterranean formation, wherein the target seismic volume comprises the seismic data;
wherein applying the machine learning model to the target seismic volume comprises:
placing, for an iteration of a plurality of iterations, a sliding window over a seismic section of the seismic data to generate a partial result, the partial result comprising at least one selected from a group consisting of a fault score and a fault detection mask, and

wherein the machine learning model comprises at least one selected from a group consisting of:
a first convolution neural network (ConvNet) for generating the fault score; and
a second ConvNet, based on at least one intermediate layer of the first ConvNet, for generating the fault detection mask.

7. The method of claim 6 , wherein the at least one selected from a group consisting of the fault score and the fault detection mask is generated, during the iteration and as the partial result, for at least one selected from a group consisting of the sliding window and a prediction area of a plurality of prediction areas of the sliding window.

8. A non-transitory computer readable medium storing instructions for causing a computer processor to carry out the method according to any of claims 1 - 7 .

9. A system for detecting an unknown fault in a target seismic volume, comprising:
a computer processor; and
memory storing instructions executed by the computer processor, wherein the instructions comprise functionality to:
generate a plurality of patches from a training seismic volume that is separate from the target seismic volume, wherein a patch of the plurality of patches comprises a set of training areas, and wherein the plurality of patches comprises contiguous portions of multiple pixels of the training seismic volume;
generate a label for assigning to the patch, wherein the label represents a subset, of the set of training areas, intersected by a known fault specified by a user in the training seismic volume;
train, during a training phase and based at least on the label and the training seismic volume, a machine learning model; and
generate, by applying the machine learning model to the target seismic volume during a prediction phase subsequent to the training phase, a result to identify the unknown fault in the target seismic volume;

a display device for rendering the result to identify the unknown fault; and
a repository for storing the plurality of patches and the set of training areas during the training phase.

10. The system of claim 9 , the instructions further comprising functionality to:
obtain seismic data corresponding to a subterranean formation, wherein the target seismic volume comprises the seismic data;
analyze the seismic data using the machine learning model to generate the result; and
determine, based on the result, a location of the unknown fault in the subterranean formation.

11. The system of claim 9 , the instructions further comprising functionality to:
extract one or more training seismic sections from the training seismic volume wherein the plurality of patches are generated from the one or more training seismic sections,
wherein the known fault is specified in the training seismic volume by the user.

12. The system of claim 9 , wherein generating the label comprises:
identifying a pre-determined label of each training area in the set of training areas; and
including the pre-determined label of each training area in the subset into the label of the patch,
wherein said each training area in the subset comprises a plurality of pixels of the training seismic volume that are intersected by the known fault.

13. The system of claim 9 , the instructions further comprising functionality to:
obtain seismic data corresponding to a subterranean formation, wherein the target seismic volume comprises the seismic data;
analyze the seismic data using the machine learning model to generate the result; and
determine, based on the result, a location of the unknown fault in the subterranean formation,
wherein applying the machine learning model to the target seismic volume comprises:
placing, for an iteration of a plurality of iterations, a sliding window over a seismic section of the seismic data,
wherein the machine learning model is applied, during the iteration and using the sliding window, to the seismic section to generate a partial result, and

wherein generating the result comprises:
aggregating a plurality of partial results generated during the plurality of iterations to generate a heat map,
wherein the result comprises the heat map.

14. The

Assignees

• Schlumberger Technology Corp

Inventors

• Salman Nader
• Kvamme Pål Kjetil

Classifications

• G06N3/045

  Combinations of networks · CPC title

• G06N3/0464

  Convolutional networks [CNN, ConvNet] · CPC title

• G06N3/09

  Supervised learning · CPC title

• G01V1/52

  Structural details · CPC title

• G01V2210/61

  Analysis by combining or comparing a seismic data set with other data · CPC title