clara yoon, karianne bergen, ossian o’reilly, fantine huot ... · copyleaks – detect plagiarism...
TRANSCRIPT
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, 2017, Cargese
0 2 4 6 8 10−0.4
−0.3
−0.2
−0.1
0
0.1
0.2
0.3
0.4
Time (s)
Ampl
itude
Normalized time window, start at 1266.95 s
Time (s)
Freq
uenc
y (H
z)
log10(|spectral image|), window #1267
0 2 4 6 8 100
1
2
3
4
5
6
7
8
9
10
−5
−4
−3
−2
−1
0
1
2
3
4
5
wavelet transform x index
wav
elet
tran
sfor
m y
inde
x
log10(|Haar transform|), window #1267
0 10 20 30 40 50 600
5
10
15
20
25
30
−5
−4
−3
−2
−1
0
1
2
3
4
5
wavelet transform x index
wav
elet
tran
sfor
m y
inde
x
Sign of top wavelet coefficients, window #1267
0 10 20 30 40 50 600
5
10
15
20
25
30
−1
0
1
fingerprint x index
finge
rprin
t y in
dex
Binary fingerprints, window #1267
0 20 40 600
10
20
30
40
50
60
0
1
Earthquake Monitoring Across Scales(meters)
Global(107)
Local(104-5)
Reservoir(103-4)
Regional(105-6)
Lab(<100)
DecreasingArrayAperture
IncreasingSensorDensity
IncreasingFrequencyCommonGoal:
-detect-locate-characterize
earthquakesascompletelyandasaccuratelyaspossible.
Mines(101-3)
LongDuraOon(Large-T)Yearsofcon+nuouswaveforms
BigNetworks(Large-N)1000sofsensors
WeneedscalablealgorithmstoextractusefulinformaOonfromthesemassivedatavolumes
Seismologyhaslotsofdata…
Standard Approach to Detection/Location
(1)DetecOon(STA/LTA)(2)AssociaOon(3)LocaOon(4)CharacterizaOon
EarleandShearer[1994]
Standard approach works well when…
Eventsarerecordedat>3staOonsEventsareimpulsiveEventsdon’toverlap
…weakeventswithemergentarrivals(likeLFEs)
Standard approach works less well for …
KatsumataandKamaya[2003]
…smalleventswithtoofewarrivalstolocate
Standard approach works less well for …
…OverlappingEventsduringintenseacOvity
Standard approach works less well for …
Detection SensitivityCo
mpu
tatio
nal E
ffici
ency
STA/LTA
General Applicability
HIGH
LOWInsensiOvetoweakand/ornon-impulsivesignals,needmulOplestaOons
HIGH
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.
38 Repeats of Earthquake on the Calaveras Fault
Adjacent Earthquakes on the Calaveras FaultEarthstructureisessenOallyconstantAdjacentearthquakeshavesimilarwaveforms.Candetectearthquakesbysearchingforsimilarwaveforms
Shellyetal.[2007]
Template Matching
Template-based detection is powerful (few Type II errors)
LFEsplantedinrealdataatsnrof0.134/36aredetected
Shellyetal.[2007]
Detection Sensitivity
Com
puta
tiona
l Effi
cien
cy
STA/LTAA
General Applicability
STemplate Matching
LOWNeedtemplatewaveformapriori
HIGH
HIGH
Informed Similarity Search
Exhaustive Search for Similar Waveforms“Autocorrela+on”-Uninformedsearchforsimilarsignal–detecteventsbycrosscorrela+ngallwindowpairs.
Brownetal.[2008]
Autocorrelation for LFE Detection
Aguiaretal.[2017]
Aguiaretal.[2017]
Detection SensitivityCo
mpu
tatio
nal E
ffici
encyTemplate mmpm
Matchingat g STA/LTAA
General Applicability
S
AutocorrelationHIGHHIGH
LOWNaïve Uninformed Similarity Search
Detection SensitivityC
omp
utat
iona
l Effi
cien
cyGeneral Applicability
New approach: FAST
mplate mplate TempmmpmmpmpmmmmmmmmmtchingMaatt g A/LTASTAAS AAS
AutocorrelationutocorrelatAutocorrrelatatocorutoAutocorrrelatirrelatir on
HIGHHIGH
HIGH
Shazam–idenOfysongsfromasampleofrecording.
Soundhound–idenOfysongsfromsinging(seriously).
TinEye–Searchthewebforthesourceofaknownimage.
YouTube–detectcopyrightinfringement
CopyLeaks–detectplagiarism
Altavista–removeduplicatewebpagesfromsearchresults
Some Big Data Technologies for Similarity Search
FAST (Fingerprinting And Similarity Thresholding)
Data
Preprocessing
FeatureExtrac+on
DatabaseGenera+on&Search
Post-processing
Detec+onResults AB
1. DataRepresenta+on
2. Fastapproximatesimilaritysearch
Fingerprinting
• “Fingerprint” waveform with sparse, diagnostic description• Store fingerprints in database and search it efficiently
WaveformDataCompression
BinaryFingerprint
DataCompression
ClaraYoon Fingerprint
Step 1: Time Series to Spectrogram
Step 2: Spectrogram to Spectral ImagesTofindshortdura+onevents,dividespectrogramintooverlappingspectralimages
Time (s)
Freq
uenc
y (Hz
)
log10(|spectral image|), window #1267
0 2 4 6 8 100
2
4
6
8
10
−5
0
5
Time (s)
Freq
uenc
y (Hz
)
log10(|spectral image|), window #1629
0 2 4 6 8 100
2
4
6
8
10
−5
0
5
Step 3: Spectral Image to its Wavelet Transform
Goal:compressnonsta+onaryseismicsignal– Compute2Ddiscretewavelettransform(Haarbasis)ofspectralimagetogetwaveletcoefficients
Time (s)
Freq
uenc
y (Hz
)
log10(|spectral image|), window #1267
0 2 4 6 8 100
2
4
6
8
10
−5
0
5
wavelet transform x index
wave
let tra
nsfor
m y i
ndex
log10(|Haar transform|), window #1267
0 20 40 600
5
10
15
20
25
30
−5
0
5
Step 4: Wavelet Transform to Top Coefficients
Keydiscrimina+vefeaturesareconcentratedinafewwaveletcoefficientswithhighestdevia+on
– Keeponlysign(+or-)ofthesecoefficients,setrestto0
Datacompression,robusttonoise
wavelet transform x index
wave
let tra
nsfor
m y i
ndex
log10(|Haar transform|), window #1267
0 20 40 600
5
10
15
20
25
30
−5
0
5
wavelet transform x index
wave
let tra
nsfor
m y i
ndex
Sign of top wavelet coefficients, window #1267
0 20 40 600
5
10
15
20
25
30
−1
0
1
Step 5: Top Coefficients into a Binary Fingerprint
FingerprintmustbecompactandsparsetostoreindatabaseConverttopcoefficientstoabinarysequenceof0’s,1’s
• Nega+ve:01,Zero:00,Posi+ve:10
wavelet transform x index
wave
let tra
nsfor
m y i
ndex
Sign of top wavelet coefficients, window #1267
0 20 40 600
5
10
15
20
25
30
−1
0
1
fingerprint x index
finge
rprint
y ind
ex
Binary fingerprints, window #1267
0 20 40 600
10
20
30
40
50
60
0
1
Jaccard SimilarityHowsimilarare2binaryfingerprints?
Jaccardsimilarity:“resemblance”
J A,B( ) =A∩BA∪B
A B
J A,B( ) =A∩BA∪B
A B
0 5 10−0.5
0
0.5
Time (s)
Ampli
tude
Normalized time window, start 1629 s
Similar Waveforms èSimilar Fingerprints
0 5 10−0.5
0
0.5
Time (s)
Ampli
tude
Normalized time window, start 1266.95 s
fingerprint x index
finge
rprint
y ind
ex
Binary fingerprints, window #1267
0 20 40 600
10
20
30
40
50
60
0
1
fingerprint x index
finge
rprint
y ind
ex
Binary fingerprints, window #1629
0 20 40 600
10
20
30
40
50
60
0
1
Fingerprints Should be Discriminative
0 5 10−0.5
0
0.5
Time (s)
Ampli
tude
Normalized time window, start 1266.95 s
0 5 10−0.5
0
0.5
Time (s)
Ampli
tude
Normalized time window, start 2876 s
fingerprint x index
finge
rprint
y ind
ex
Binary fingerprints, window #2876
0 20 40 600
10
20
30
40
50
60
0
1
fingerprint x index
finge
rprint
y ind
ex
Binary fingerprints, window #1267
0 20 40 600
10
20
30
40
50
60
0
1
Names are a compact “Fingerprint”
Names are compact, but not discriminative.
How to select “discriminative” coefficients?
Red represents intersection of fingerprints for samples 1 and 2.
0.266 0.117 0.040
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-offCurves
FAST Workflow
Locality Sensitive Hashing groups similar fingerprints drawn from a high-dimensional space with high probability
Waveformsearchquery Database
sMatch!
Yoonetal.(2015)
Min-Hash uses multiple random hash functions to map a binary fingerprint to a single integer.
The probability of two finger- prints 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.
Why is FAST Fast for Large T?
Ignores>1011pairs
Detects<103pairs
Outputs<105pairs
A
B
1 week 2 weeks 1 month 3 months 6 months
3 days 1 week 2 weeks 1 month 3 months 6 months1 day
1 day
1 week
1 hour
1 year
20 years
Yoonetal.(2015)
Why is FAST Fast for Large T?
Ignores>1011pairs
Detects<103pairs
Outputs<105pairs
Yoonetal.(2015)
80km
Guy-Greenbrier Sequence in Arkansas
HydraulicS+mula+on(Fracking)wellsuseastagedinjec+onof
fluidtoincreasepermeabilityandaccesshydrocarbons.
Deepdisposalwellsinjectproducedwater,or
frackingflowbackwater,togetridofit.
Guy-Greenbrier, Arkansas Sequence
92.5˚W 92.4˚W 92.3˚W 92.2˚W 92.1˚W35.2˚N
35.3˚N
35.4˚N
92.5˚W 92.4˚W 92.3˚W 92.2˚W 92.1˚W35.2˚N
35.3˚N
35.4˚N
0123456789
10
Earth
quak
e D
epth
km
WHAR
ARK1
ARK2
M=4.7
Well 1
Well 2
Well 3
Well 4
Well 5
Well 6
Well 8
Guy
Greenbrier0 10
km
N
-Startedinjec+ng2010-07-07
-Startedinjec+ng2010-08-16
Injec3ondata:Horton(2012),SRL
Swarmdura+on:July2010–June2011
2011-02-27
Catalogevents(Seismik):2010-06-01to2010-09-01
Wastewaterinjec+onwell
Town
Seismicsta+on
Catalogevents(ANSS):2010-09-01to2011-10-31
3 Months Guy-Greenbrier Induced Seismicity
3 Quarry Blasts
43/24
Quarry Image, 2009-07-23
44/24
Quarry Image, 2010-09-15
Mostlikelyloca+onofquarryblastson2010-06-24,2010-07-02,2010-08-10:(35.2928,-92.3973,0kmdepth)
2 3 4 5 6 7Velocity (km/s)
0
5
10
15
20
Dept
h (k
m)
1D Velocity Model
Vp, Ogwari 2016Vs, Ogwari 2016Vp, New Model from QuarryVs, New Model from Quarry
ImprovedPandS-wavevelocitystructurebasedonknownquarryblastsusingVelest.Sparsenetwork–(3)3-componentstaOonshypo-DDusingPandS700maposterioriadjustment.
92.5˚W 92.4˚W 92.3˚W 92.2˚W 92.1˚W35.2˚N
35.3˚N
35.4˚N
92.5˚W 92.4˚W 92.3˚W 92.2˚W 92.1˚W35.2˚N
35.3˚N
35.4˚N
0123456789
10
Depth
km
WHAR
ARK1
ARK2
Quarry
Well 1
Well 2
Well 3
Well 4
Well 5
Well 6
Well 8
0 10
km
N
Group Earthquakes into Clusters �(-1.1 < M < 1.8)
46Earthquakes2010-06-01to2010-09-01
Wastewaterinjec+onwellProduc+onwells+mulated2010-06-01to2010-09-01
92.5˚W 92.4˚W 92.3˚W 92.2˚W 92.1˚W35.2˚N
35.3˚N
35.4˚N
92.5˚W 92.4˚W 92.3˚W 92.2˚W 92.1˚W35.2˚N
35.3˚N
35.4˚N
0123456789
10
Depth
km
WHAR
ARK1
ARK2
Quarry
Well 1
Well 2
Well 3
Well 4
Well 5
Well 6
Well 8
0 10
km
N
47
C#1
C#2
C#3
C#4
C#5C#6
C#7
C#8
C#9
C#10
C#11C#12
C#13
C#14C#15
C#16Earthquakes2010-06-01to2010-09-01
Wastewaterinjec+onwellProduc+onwells+mulated2010-06-01to2010-09-01
C#18
Group Earthquakes into Clusters �(-1.1 < M < 1.8)
S+mula+onatnearbywell
Cluster #1: 3143 events �(667 located + 2525 assigned)
Further Evidence for Earthquakes �Induced by Hydraulic Stimulation
92.3˚W92.3˚W
0123456789
10
Dep
th
kmARK2
Well 1
0 1 2
km
Composite Focal MechanismN60°EShmax(HurdandZoback,2012]
If double-couple, plane of seismicity is right-lateral – contrary to stress. More likely that first motions reflect combined tensile & shear failure.
ConsistentwithhydraulicdiffusivityofD≈1m2/s
Conclusions I
BothwastewaterinjecOonandhydraulicsOmulaOonappeartotriggerearthquakes–probablysomenaturalearthquakesaswell.Itischallengingtodisentangledifferentinfluences–requiresgooddata(bothseismicandinjecOon).PrecisionseismologyisapowerfultooltoprovideaclearerpictureofinducedseismicityandthenucleaOonprocesstotheextentitisexpressedinseismicity.
Conclusions II
Seismologyhas:• LongduraOondata(Large-T)• Bignetworks(Large-N)
èFASTalgorithmenablesdata-drivendiscovery
Moredata
Morememory
MorecompuOngpower
Needbemeralgorithms
Now Future
Scale of Effort
Yoon et al. (2015) (2017a) (2017b)
Progress on FAST for Large-T Problems
• 140xFASTerthanoriginal.
• Reducedmemoryrequirements.
• ReducedfalsedetecOonrate.
• Improvedpost-processing.
• Workingtowardspublicrelease.
FAST for a Decade of Continuous Data
Yoonetal.(2017b)
FAST over a Network
BergenandBeroza(2017)
Network FAST for Iquique
~580newdetecOonsin17daysbeforethemainshock.Canbeusedastemplatestoincreasethatnumber.
Machine Learning for Earthquake Detection
Huotetal.(2017)
Labeleddataasinputtoneuralnetwork(mostofwhatwerecordisnoise)
Huotetal.(2017)
99.5%accuracywhentrained,validatedandtestedononestaOon.Accuracydropsto98.2%,withmulOplestaOonsbutwithonlyalimiteddataset.
ML for Earthquake Detection
Scale of Seismic Observations
Nakata (2017)
Merci