Earthquakes: Nucleation, Triggering, Rupture, and Relationships to Aseismic Processes October 3, 2017, Cargese. ( Haar transform ), window #1267

Transcription

1 The FAST Method for Earthquake Detection: Application to Seismicity During the Initial Stages of the Guy- Greenbrier, Arkansas, Earthquake Sequence Clara Yoon, Karianne Bergen, Ossian O Reilly, Fantine Huot, Bill Ellsworth, Greg Beroza (Stanford University) Yihe Huang (University of Michigan) Earthquakes: Nucleation, Triggering, Rupture, and Relationships to Aseismic Processes October 3, 217, Cargese Amplitude Normalized time window, start at s Time (s) Frequency (Hz) log 1 ( spectral image ), window # Time (s) wavelet transform y index log 1 ( Haar transform ), window # wavelet transform x index wavelet transform y index Sign of top wavelet coefficients, window # wavelet transform x index 1 fingerprint y index Binary fingerprints, window # fingerprint x index 1

2 Earthquake Monitoring Across Scales (meters) Global (1 7 ) Decreasing Array Aperture Common Goal: - detect - locate - characterize earthquakes as completely and as accurately as possible. Regional (1 5-6 ) Local (1 4-5 ) Reservoir (1 3-4 ) Increasing Sensor Density Mines (1 1-3 ) Increasing Frequency Lab (< 1 )

3 Seismology has lots of data Big Networks (Large-N) 1s of sensors Long DuraOon (Large-T) Years of con+nuous waveforms We need scalable algorithms to extract useful informaoon from these massive data volumes

4 Standard Approach to Detection/Location (1) DetecOon (STA/LTA) (2) AssociaOon (3) LocaOon (4) CharacterizaOon Earle and Shearer [1994]

5 Standard approach works well when Events are recorded at > 3 staoons Events are impulsive Events don t overlap

6 Standard approach works less well for weak events with emergent arrivals (like LFEs) Katsumata and Kamaya [23]

7 Standard approach works less well for small events with too few arrivals to locate

8 Standard approach works less well for Overlapping Events during intense acovity

9 Computational Efficiency HIGH STA/LTA LOW Detection Sensitivity General Applicability InsensiOve to weak and/or nonimpulsive signals, need mulople staoons HIGH

10 38 Repeats of Earthquake on the Calaveras Fault Slip occurring at different times in the same place, generates identical seismograms. We can look for a repeating signal from a repeating source, but most sources don t exactly repeat.

11 Adjacent Earthquakes on the Calaveras Fault Earth structure is essenoally constant Adjacent earthquakes have similar waveforms. Can detect earthquakes by searching for similar waveforms

12 Shelly et al. [27] Template Matching

13 Template-based detection is powerful (few Type II errors) LFEs planted in real data at snr of.1 34/36 are detected Shelly et al. [27]

14 Computational Efficiency Informed Similarity Search Template Matching HIGH STA/LTA HIGH Detection Sensitivity General Applicability LOW Need template waveform a priori

15 Exhaustive Search for Similar Waveforms Autocorrela+on - Uninformed search for similar signal detect events by cross correla+ng all window pairs. Brown et al. [28]

16 Autocorrelation for LFE Detection Aguiar et al. [217]

17 Aguiar et al. [217]

18 Computational Efficiency Naïve Uninformed Similarity Search LOW Template Matching STA/LTA HIGH Detection Sensitivity General Applicability Autocorrelation HIGH

19 Computational Efficiency Temp mplate Matching HIGH STA A/LTA New approach: FAST HIGH Detection Sensitivity General Applicability Autocorrelation on HIGH

20 Shazam idenofy songs from a sample of recording. Soundhound idenofy songs from singing (seriously). Some Big Data Technologies for Similarity Search CopyLeaks detect plagiarism TinEye Search the web for the source of a known image. YouTube detect copyright infringement Altavista remove duplicate web pages from search results

21 FAST (Fingerprinting And Similarity Thresholding) Data Preprocessing Feature Extrac+on Database Genera+on & Search Post-processing Detec+on Results A 1. Data Representa+on 2. Fast approximate similarity search B

22 Fingerprinting Clara Yoon Fingerprint Data Compression Waveform Data Compression Binary Fingerprint Fingerprint waveform with sparse, diagnostic description Store fingerprints in database and search it efficiently

23 Step 1: Time Series to Spectrogram

24 Step 2: Spectrogram to Spectral Images To find short dura+on events, divide spectrogram into overlapping spectral images log 1 ( spectral image ), window #1267 log 1 ( spectral image ), window # Frequency (Hz) 6 4 Frequency (Hz) Time (s) Time (s) 8 1 5

25 1 Step 3: Spectral Image to its Wavelet Transform log ( spectral image ), window #1267 log 1 1 ( Haar transform ), window # Frequency (Hz) wavelet transform y index Time (s) wavelet transform x index Goal: compress nonsta+onary seismic signal Compute 2D discrete wavelet transform (Haar basis) of spectral image to get wavelet coefficients 5

26 3 Step 4: Wavelet Transform to Top Coefficients log 1 ( Haar transform ), window #1267 Sign of top wavelet coefficients, window # wavelet transform y index wavelet transform y index wavelet transform x index Key discrimina+ve features are concentrated in a few wavelet coefficients with highest devia+on Keep only sign (+ or -) of these coefficients, set rest to Data compression, robust to noise wavelet transform x index 1

27 Step 5: Top Coefficients into a Binary Fingerprint Sign of top wavelet coefficients, window #1267 Binary fingerprints, window # wavelet transform y index fingerprint y index wavelet transform x index Fingerprint must be compact and sparse to store in database Convert top coefficients to a binary sequence of s, 1 s Nega+ve: 1, Zero:, Posi+ve: fingerprint x index

28 Jaccard Similarity How similar are 2 binary fingerprints? A B Jaccard similarity: resemblance ( ) = A B J A, B A B A B

29 Similar Waveforms è Similar Fingerprints Normalized time window, start s.5 Binary fingerprints, window # Amplitude fingerprint y index Time (s) Normalized time window, start 1629 s fingerprint x index 6 Binary fingerprints, window # Amplitude fingerprint y index Time (s) fingerprint x index

30 Fingerprints Should be Discriminative Normalized time window, start s.5 Binary fingerprints, window # Amplitude fingerprint y index Time (s) Normalized time window, start 2876 s fingerprint x index 6 Binary fingerprints, window # Amplitude fingerprint y index Time (s) fingerprint x index

31 Names are a compact Fingerprint Names are compact, but not discriminative.

32 How to select discriminative coefficients? Red represents intersection of fingerprints for samples 1 and 2.

33 True Detections vs. False Positives Don t choose largest coefficients, choose those on the tails of a distribution. Suppresses false detections of persistent noise, but maintains high accuracy for relatively rare earthquake signals. Trade-off Curves

34 FAST Workflow Min-Hash uses multiple random hash functions to map a binary fingerprint to a single integer. The probability of two fingerprints A and B mapping to the same integer is equal to their Jaccard similarity. Min-Hash reduces dimensionality while preserving the similarity between A and B in a probabilistic manner. Locality Sensitive Hashing groups similar fingerprints drawn from a high-dimensional space with high probability Waveform search query s Match! Database Yoon et al. (215)

35 Why is FAST Fast for Large T? B 2 years Ignores > 1 11 pairs Outputs < 1 5 pairs Detects < 1 3 pairs 1 year 1 week 1 day 1 hour 1 week 2 weeks 1 month 3 months 6 months Yoon et al. (215)

36 Why is FAST Fast for Large T? Ignores > 1 11 pairs Outputs < 1 5 pairs Detects < 1 3 pairs Yoon et al. (215)

37 Guy-Greenbrier Sequence in Arkansas 8 km

38 Deep disposal wells inject produced water, or fracking flowback water, to get rid of it. Hydraulic S+mula+on (Fracking) wells use a staged injec+on of fluid to increase permeability and access hydrocarbons.

39 Guy-Greenbrier, Arkansas Sequence 35.4 N Swarm dura+on: July 21 June 211 km 1 Well 8 9 ARK2 8 Well 4 7 Well 6 Guy 6 Well 1 - Started injec+ng N 5 WHAR 4 3 ARK1 Well Started injec+ng Well 2 M= km Greenbrier Well N 92.5 W 92.4 W 92.3 W 92.2 W 92.1 W N Earthquake Depth Wastewater injec+on well Seismic sta+on Town Catalog events (Seismik): to Catalog events (ANSS): to Injec3on data: Horton (212), SRL

40 3 Months Guy-Greenbrier Induced Seismicity

41

42 3 Quarry Blasts

43 Quarry Image, /24

44 Quarry Image, Most likely loca+on of quarry blasts on , , : ( , , km depth) 44/24

45 1D Velocity Model Depth (km) Vp, Ogwari 216 Vs, Ogwari 216 Vp, New Model from Quarry Vs, New Model from Quarry Velocity (km/s) Improved P and S-wave velocity structure based on known quarry blasts using Velest. Sparse network (3) 3- component staoons hypo-dd using P and S 7 m a posteriori adjustment.

46 35.4 N Group Earthquakes into Clusters (-1.1 < M < 1.8) 35.3 N Well 8 Well 6 Quarry Well 2 ARK1 Well 1 Well 5 ARK2 WHAR Well 4 Wastewater injec+on well km Produc+on well s+mulated to Well N Earthquakes to W 92.4 W 92.3 W 92.2 W 92.1 W N km Depth

47 35.4 N 35.3 N Well 8 Group Earthquakes into Clusters Well 6 C#14 Quarry C#15 Well 2 (-1.1 < M < 1.8) C#9 C#8 C#4 ARK1 C#7 Well 1 Well 5 C#1 C#2 ARK2 C#3 WHAR C#5 Well 4 C#12 C#11 C#18 C#6 C#13 Wastewater injec+on well km Produc+on well s+mulated C#16 C# to Well N Earthquakes to W 92.4 W 92.3 W 92.2 W 92.1 W N km Depth

48 S+mula+on at nearby well

49 Cluster #1: 3143 events (667 located assigned)

50 Further Evidence for Earthquakes Induced by Hydraulic Stimulation Well 1 ARK2 km 1 2 km Depth 92.3 W

51

52 Composite Focal Mechanism N6 E Shmax (Hurd and Zoback, 212] If double-couple, plane of seismicity is right-lateral contrary to stress. More likely that first motions reflect combined tensile & shear failure.

53 Consistent with hydraulic diffusivity of D 1 m 2 /s

54 Conclusions I Both wastewater injecoon and hydraulic somulaoon appear to trigger earthquakes probably some natural earthquakes as well. It is challenging to disentangle different influences requires good data (both seismic and injecoon). Precision seismology is a powerful tool to provide a clearer picture of induced seismicity and the nucleaoon process to the extent it is expressed in seismicity.

55 Now Seismology has: Long duraoon data (Large-T) Big networks (Large-N) Conclusions II More data Future è FAST algorithm enables data-driven discovery More memory Need bemer algorithms More compuong power

56 Progress on FAST for Large-T Problems 14x FASTer than original. Scale of Effort Reduced memory requirements. Reduced false detecoon rate. Improved post-processing. Working towards public release. Yoon et al. (215) (217a) (217b)

57 FAST for a Decade of Continuous Data Yoon et al. (217b)

58 FAST over a Network Bergen and Beroza (217)

59 Network FAST for Iquique ~58 new detecoons in 17 days before the mainshock. Can be used as templates to increase that number.

60 Machine Learning for Earthquake Detection Labeled data as input to neural network (most of what we record is noise) Huot et al. (217)

61 ML for Earthquake Detection 99.5% accuracy when trained, validated and tested on one staoon. Accuracy drops to 98.2%, with mulople staoons but with only a limited data set. Huot et al. (217)

62 Scale of Seismic Observations Nakata (217)

63 Merci