Event detection and de-noising method for passive seismic data
View Patent ↗An apparatus, a method, and a non-transitory computer readable medium for event detection of passive seismic data are disclosed. The apparatus includes processing circuitry extracts features from the passive seismic data based on a backbone subnetwork of a residual deep neural network. The processing circuitry generates bounding box proposals for a region of interest (ROI) in the passive seismic data based on the extracted features being input to a region proposal network of the residual deep neural network. The processing circuitry classifies the bounding box proposals into two groups. Each bounding box proposal in a first group indicates that a corresponding seismic signal presents in the ROI. Each bounding box proposal in a second group indicates that no seismic signal presents in the ROI. The processing circuitry determines at least one seismic signal in the ROI from the first group of bounding box proposals.
1. A method for event detection of passive seismic data, the method comprising:
extracting features from the passive seismic data based on a backbone subnetwork of a residual deep neural network;
generating bounding box proposals for a region of interest (ROI) in the passive seismic data based on the extracted features being input to a region proposal network of the residual deep neural network;
adjusting parameters of the backbone subnetwork based on the extracted features;
classifying the bounding box proposals into two groups, each bounding box proposal in a first group indicating that a corresponding seismic signal presents in the ROI, each bounding box proposal in a second group indicating that no seismic signal presents in the ROI; and
determining at least one seismic signal in the ROI from the first group of bounding box proposals.
2. The method of claim 1 , further comprising:
flipping each bounding box proposal in the first group in one of a horizontal direction or a vertical direction along with the respective bounding box proposal.
3. The method of claim 1 , further comprising:
de-noising the at least one seismic signal based on an infinite impulse response (IIR) Wiener filter based de-noising method.
4. The method of claim 1 , further comprising:
increasing a size of each bounding box proposal in the first group along a diagonal associated with a seismic signal detected inside the respective bounding box proposal, an aspect ratio of each bounding box proposal being kept as constant during the increasing the size of the respective bounding box proposal.
5. The method of claim 1 , further comprising:
generating synthetic seismic data based on a Ricker wavelet;
training the residual deep neural network based on a portion of the synthetic seismic data; and
testing the residual deep neural network based on a remaining portion of the synthetic seismic data.
6. The method of claim 5 , wherein the testing comprises:
adding random noise to the remaining portion of the synthetic seismic data to generate noisy synthetic seismic data; and
testing the residual deep neural network based on the noisy synthetic seismic data.
7. A method for event detection of passive seismic data, the method comprising:
extracting features from the passive seismic data based on a backbone subnetwork of a residual deep neural network;
generating bounding box proposals for a region of interest (ROI) in the passive seismic data based on the extracted features being input to a region proposal network of the residual deep neural network;
classifying the bounding box proposals into two groups, each bounding box proposal in a first group indicating that a corresponding seismic signal presents in the ROI, each bounding box proposal in a second group indicating that no seismic signal presents in the ROI;
flipping each bounding box proposal in the first group in one of a horizontal direction or a vertical direction along with the respective bounding box proposal; and
determining at least one seismic signal in the ROI from the first group of bounding box proposals.
8. The method of claim 7 , further comprising:
de-noising the at least one seismic signal based on an infinite impulse response (IIR) Wiener filter based de-noising method.
9. The method of claim 7 , further comprising:
increasing a size of each bounding box proposal in the first group along a diagonal associated with a seismic signal detected inside the respective bounding box proposal, an aspect ratio of each bounding box proposal being kept as constant during the increasing the size of the respective bounding box proposal.
10. The method of claim 7 , further comprising:
generating synthetic seismic data based on a Ricker wavelet;
training the residual deep neural network based on a portion of the synthetic seismic data; and
testing the residual deep neural network based on a remaining portion of the synthetic seismic data.
11. The method of claim 10 , wherein the testing comprises:
adding random noise to the remaining portion of the synthetic seismic data to generate noisy synthetic seismic data; and
testing the residual deep neural network based on the noisy synthetic seismic data.
12. A method for event detection of passive seismic data, the method comprising:
extracting features from the passive seismic data based on a backbone subnetwork of a residual deep neural network;
generating bounding box proposals for a region of interest (ROI) in the passive seismic data based on the extracted features being input to a region proposal network of the residual deep neural network;
classifying the bounding box proposals into two groups, each bounding box proposal in a first group indicating that a corresponding seismic signal presents in the ROI, each bounding box proposal in a second group indicating that no seismic signal presents in the ROI;
increasing a size of each bounding box proposal in the first group along a diagonal associated with a seismic signal detected inside the respective bounding box proposal, an aspect ratio of each bounding box proposal being kept as constant during the increasing the size of the respective bounding box proposal; and
determining at least one seismic signal in the ROI from the first group of bounding box proposals.
13. The method of claim 12 , further comprising:
adjusting parameters of the backbone subnetwork based on the extracted features.
14. The method of claim 12 , further comprising:
de-noising the at least one seismic signal based on an infinite impulse response (IIR) Wiener filter based de-noising method.
15. The method of claim 12 , further comprising:
generating synthetic seismic data based on a Ricker wavelet;
training the residual deep neural network based on a portion of the synthetic seismic data; and
testing the residual deep neural network based on a remaining portion of the synthetic seismic data.
16. The method of claim 15 , wherein the testing comprises:
adding random noise to the remaining portion of the synthetic seismic data to generate noisy synthetic seismic data; and
testing the residual deep neural network based on the noisy synthetic seismic data.