volcanic earthquake timing using wireless sensor · pdf filevolcanic earthquake timing using...

Volcanic Earthquake Timing

Using Wireless Sensor Networks

Guojin Liu1,2 Rui Tan2,3 Ruogu Zhou2 Guoliang Xing2

Wen-Zhan Song4 Jonathan M. Lees5

1Chongqing University, P.R. China2Michigan State University, USA

3Advanced Digital Science Center, Illinois at Singapore4Georgia State University, USA

5University of North Carolina at Chapel Hill, USA

Volcano Hazards

• 7% world population live near active volcanoes

• 20 - 30 explosive eruptions/year

Eruption in Chile, 6/4, 2011

$68 M instant damage, $2.4 B future relief.www.boston.com/bigpicture/2011/06/volcano_erupts_in_chile.html

Eruptions in Iceland 2010

A week-long airspace closure

[Wikipedia]

2

Volcano Monitoring

• Seismic activity monitoring– Earthquake localization, tomography, early warning etc.

• Traditional seismometer– Expensive (~$10K/unit), difficult to install & retrieve

– Only ~10 nodes installed for most threatening volcanoes!

Photo credit: USGS, http://volcanoes.usgs.gov/activity/methods/ 3

Sensor Networks for Volcano Monitoring

• Sensor systems for volcano monitoring

– Harvard , OASIS@GSU, VolcanoSRI@GSU/MSU/UNC

– Raw data collection@100Hz & centralized analysis

– Short lifetime (~1 week)

• In-network earthquake detection [Tan 2010]

– Distributed seismic signal processing

– 83% energy reduction from raw data collection

OASIS node

Harvardnode

4

Earthquake Timing

Node 05

Node 04

Node 06

0 1 2 3

P-phase

• Key to localization, seismic tomography, etc.– Usually done manually, automation is expensive

Time (second)0 1 2 3

Earthquake timingSource localization

Seismic tomography

5

Earthquake Timing

0 1 2 3 4

Node 09

Node 10?

• Key to localization, seismic tomography, etc.– Usually done manually, automation is expensive

• In-situ P-phase picking w/ limited transmission– Data intensive

– Sensors have limited compute & comm. capabilities

Source localization

Seismic tomography

0 1 2 3 4

5

Seismic Signal: Sparsity

Original @ 100Hz

Time (second)

6

Seismic Signal: Sparsity

Original @ 100Hz

Time (second)

Sparsity=0.57

%5||||

||||

2

2)( <−

s

ss kK-sparse signal:

length signalsparsity

k=

wavelet K largest

points

6

Seismic Signal: Sparsity

Original @ 100Hz

Wavelet

Time (second)

Sparsity=0.57

Sparsity=0.14

Time-frequency domain

%5||||

||||

2

2)( <−

s

ss kK-sparse signal:

length signalsparsity

k=

wavelet K largest

points

Sparsity=0.14

6

Seismic Signal: Sparsity

Original @ 100Hz

Wavelet

Time (second)

Sparsity=0.57

Sparsity=0.14

Time-frequency domain

%5||||

||||

2

2)( <−

s

ss kK-sparse signal:

length signalsparsity

k=

wavelet K largest

points

Sparsity=0.14

• Observation 1: wavelet sparsifies signal

6

Seismic Signal: Frequency-Time

4-level wavelet transform (length=1600)

thumbnail (length=100)

Low-pass band (0, 6.25Hz)

original

Time (unit: 160ms)

Low-pass band (0, 6.25Hz)

P-wave < 5Hz

7

Seismic Signal: Frequency-Time

4-level wavelet transform (length=1600)

thumbnail (length=100)

Low-pass band (0, 6.25Hz)

original

Time (unit: 160ms)

Low-pass band (0, 6.25Hz)

P-wave < 5Hz

Original (length=1600)

7

Seismic Signal: Frequency-Time

4-level wavelet transform (length=1600)

thumbnail (length=100) Rough P-phase

estimate

original

Time (unit: 160ms)

estimate

Original (length=1600)

7

Seismic Signal: Frequency-Time

4-level wavelet transform (length=1600)

thumbnail (length=100) Rough P-phase

estimate

original

Time (unit: 160ms)

estimate

• Observation 2: P-phase estimate from thumbnail

Original (length=1600)

7

Seismic Signal: Diversity

Node 1sparsity=0.1

Node 10sparsity=0.38

Earthquake 1

8

Seismic Signal: Diversity

Node 1sparsity=0.1

Node 10sparsity=0.38

Earthquake 1

Earthquake 2

Node 10sparsity=0.14

8

Seismic Signal: Diversity

Node 1sparsity=0.1

Node 10sparsity=0.38

Earthquake 1

• Observation 3: sensors have different sparsities

Earthquake 2

Node 10sparsity=0.14

8

Outline

• Problem statement

• Approach overview

• Earthquake timing algorithms

• Performance evaluation• Performance evaluation

• Conclusion

9

Approach Overview

Cluster head

10

Approach Overview

Cluster head

10

Approach Overview

preliminary picksignal sparsity

preliminary picksignal sparsity

preliminary picksignal sparsity

preliminary picksignal sparsitypreliminary pick

signal sparsity

Cluster head

• Lightweight signal processing algorithms– Signal sparsity

– Preliminary P-phase from thumbnail

10

Approach Overview

preliminary picksignal sparsity

preliminary picksignal sparsity

preliminary picksignal sparsity

preliminary picksignal sparsitypreliminary pick

signal sparsity

Cluster head

• Lightweight signal processing algorithms– Signal sparsity

– Preliminary P-phase from thumbnail

10

Approach Overview

Cluster head

• Lightweight signal processing algorithms– Signal sparsity

– Preliminary P-phase from thumbnail

• Select most informative sensors to TX

10

Approach Overview

××××0010

1001

0100

××××0010

1001

0100

0100

Cluster head

• Lightweight signal processing algorithms– Signal sparsity

– Preliminary P-phase from thumbnail

• Select most informative sensors to TX– Compressive sampling & transmission

××××0010

1001

10

Approach Overview

Signal

reconstruction

Signal

reconstruction

accurate pick

accurate pick

accurate pick

Source localization

Seismic tomography

…

• Lightweight signal processing algorithms– Signal sparsity

– Preliminary P-phase from thumbnail

• Select most informative sensors to TX– Compressive sampling & transmission

10

Outline

• Problem statement

• Approach overview

• Earthquake timing algorithms

– Pre-processing @ sensors– Pre-processing @ sensors

– Sensor selection & compressive sampling

• Performance evaluation

• Conclusion

11

Preliminary P-phase Pick4-level wavelet transform (length=1600)

thumbnail (length=100)

Time (unit: 160ms)

12

Preliminary P-phase Pick4-level wavelet transform (length=1600)

thumbnail (length=100)

preliminary pick

Time (unit: 160ms)

p

p

p beforeenergy signal

after energy signalmaxarg2picky preliminar

thumbnail

4

∈×=

12

Preliminary P-phase Pick4-level wavelet transform (length=1600)

thumbnail (length=100)

preliminary pick

Time (unit: 160ms)

p

p

p beforeenergy signal

after energy signalmaxarg2picky preliminar

thumbnail

4

∈×=

Map thumbnail domain back

to original time domain

12

Preliminary P-phase Pick4-level wavelet transform (length=1600)

thumbnail (length=100)

preliminary pick

Time (unit: 160ms)

p

p

p beforeenergy signal

after energy signalmaxarg2picky preliminar

thumbnail

4

∈×=

Map thumbnail domain back

to original time domain

• Lightweight: O(signal length)

– Suitable for resource-constrained sensors12

Outline

• Problem statement

• Approach overview

• Earthquake timing algorithms

– Pre-processing @ sensors– Pre-processing @ sensors

– Sensor selection & compressive sampling

• Performance evaluation

• Conclusion

13

Impact of Timing on Source Localization

• Source localization

– Basis for many volcano monitoring applications

– Complex non-linear inverse problem

z1z2

z

• Information-theoretic error metric

( )( )1tr

−= TE GG scaled Fisher matrix:

zi, z0

z1

z0

z3Vt1

t2

t3

),,(gray tracin 0 Vzzt ii =

14

Impact of Timing on Source Localization

• Source localization

– Basis for many volcano monitoring applications

– Complex non-linear inverse problem

z1z2

z

sensor

position

• Information-theoretic error metric

( )( )1tr

−= TE GG scaled Fisher matrix:

zi, z0

z1

z0

z3Vt1

t2

t3

),,(gray tracin 0 Vzzt ii =

14

Impact of Timing on Source Localization

• Source localization

– Basis for many volcano monitoring applications

– Complex non-linear inverse problem

z1z2

z

sensor

position

source

location

• Information-theoretic error metric

( )( )1tr

−= TE GG scaled Fisher matrix:

zi, z0

z1

z0

z3Vt1

t2

t3

),,(gray tracin 0 Vzzt ii =

14

Impact of Timing on Source Localization

• Source localization

– Basis for many volcano monitoring applications

– Complex non-linear inverse problem

z1z2

z

sensor

position

source

location

volcano

model

• Information-theoretic error metric

( )( )1tr

−= TE GG scaled Fisher matrix:

zi, z0

z1

z0

z3Vt1

t2

t3

),,(gray tracin 0 Vzzt ii =

14

Dynamic Sensor Selection

• Find a subset of sensors S to minimize E s.t.

CimcSi

i ≤⋅∑∈

)sensor ofsparsity (

15

Dynamic Sensor Selection

• Find a subset of sensors S to minimize E s.t.

CimcSi

i ≤⋅∑∈

)sensor ofsparsity (

unit TX cost

15

Dynamic Sensor Selection

• Find a subset of sensors S to minimize E s.t.

CimcSi

i ≤⋅∑∈

)sensor ofsparsity (

unit TX cost

TX volume

15

Dynamic Sensor Selection

• Find a subset of sensors S to minimize E s.t.

CimcSi

i ≤⋅∑∈

)sensor ofsparsity (

unit TX cost

TX volume cost budget

15

Dynamic Sensor Selection

• Find a subset of sensors S to minimize E s.t.

CimcSi

i ≤⋅∑∈

)sensor ofsparsity (

• Brutal-force search

– 8 seconds on Imote2 for 16 sensors

– Information gain diminishes for larger clusters

unit TX cost

TX volume cost budget

15

Compressive Sampling (CS)

• Apply CS to wavelet coefficients

=x

originalcompressedrandom matrix

m

n

n m

• Apply CS to wavelet coefficients

– Known TX volume before compression

– Unselected sensors avoid compression overhead

nm ××= sparsity5.1

16

Compressive Sampling (CS)

• Apply CS to wavelet coefficients

=x

originalcompressedrandom matrix

m

n

n m

• Apply CS to wavelet coefficients

– Known TX volume before compression

– Unselected sensors avoid compression overhead

nm ××= sparsity5.1

best trade-off b/w TX volume

and signal reconstruction error

16

Outline

• Problem statement

• Approach overview

• Earthquake timing algorithms

• Performance evaluation• Performance evaluation

– Testbed experiments

– Extensive trace-driven simulations

• Conclusion

17

Testbed Experiments

• Implementation on 12 TelosB

– Seismic data from Mt St Helens -> mote flash

– Real-time data acquisition @ 100 Hz

4

Exe

cuti

on

tim

e (

seco

nd

)

1 124 8

Sensor ID

0

1

2

3

Exe

cuti

on

tim

e (

seco

nd

)

End-to-end delay

< 3 seconds

18

Trace-driven Simulation

• Data traces from 12 sensors on Mt St Helens

• 30 significant earthquakes in 5.5 months

30

11

# o

f se

lect

ed

se

nso

rsLance

[SenSys’08]

200 400 600 200 400 6000

10

20

5

7

9

TX bound (# of pkts) TX bound (# of pkts)

Err

or

me

tric

# o

f se

lect

ed

se

nso

rs

our sensor

selection approach

[SenSys’08]

Configurable trade-off between system performance and energy consumption

19

Impact of Packet Loss

= xcompressed signal

30

Re

lati

ve r

eco

nst

ruct

ion

err

or

(%)

Lossy compression:

encodes largest

reconstructed signal

65 70 75 80 85 90 95 100Packet reception ratio (%)

0

10

20

Re

lati

ve r

eco

nst

ruct

ion

err

or

(%)

Compressive

Sensing

encodes largest

wavelet coefficients

CS is resilient to

packet loss!

20

Impact of Packet Loss

= xcompressed signal

received

30

Re

lati

ve r

eco

nst

ruct

ion

err

or

(%)

Lossy compression:

encodes largest

reconstructed signal

65 70 75 80 85 90 95 100Packet reception ratio (%)

0

10

20

Re

lati

ve r

eco

nst

ruct

ion

err

or

(%)

Compressive

Sensing

encodes largest

wavelet coefficients

CS is resilient to

packet loss!

20

Impact of Packet Loss

= xcompressed signal

received

30

Re

lati

ve r

eco

nst

ruct

ion

err

or

(%)

Lossy compression:

encodes largest

reconstructed signal

65 70 75 80 85 90 95 100Packet reception ratio (%)

0

10

20

Re

lati

ve r

eco

nst

ruct

ion

err

or

(%)

Compressive

Sensing

encodes largest

wavelet coefficients

CS is resilient to

packet loss!

20

Accuracy of Timingfine-grained pick on

original

fine-grained pick on

reconstructed

21

Accuracy of Timingfine-grained pick on

original

fine-grained pick on

reconstructed

16% data TX16% data TX0.6 km localization error

21

Conclusions

• Energy-efficient earthquake timing

– Lightweight algorithms for sensors

– Dynamic sensor selection

– Compressive sampling

• Testbed experiments

– Feasibility of our approach on motes

• Trace-driven simulations

– Accurate timing with 16% data transmitted

22

Hierarchical Network Architecture

6.7km

sensor

coordinator

cluster

STA/LTA detectorBayesian detector

0.1

0.2

0.3

sen

sor

/ co

ord

ina

tor

• Sensors

– Limited capability, large spatial coverage

• Coordinators

– Powerful, limited number

500 nodes on Tungurahua, Ecuador, 2015

[VolcanoSRI project]

0

sen

sor

/ co

ord

ina

tor

10 100 200

# earthquakes per day

MCU & radio energy ratio

TelosB vs. Imote2

23

Earthquake Source Localization

10

15

20

25

% o

f tr

an

smit

ted

da

ta

0.4

0.6

0.8

Sou

rce

loca

liza

tio

n e

rro

r (k

m)

Packet reception ratio

85%

180 220 260 300 360 220 260 300 3600

5

10

# of packets

% o

f tr

an

smit

ted

da

ta

0

0.2

Sou

rce

loca

liza

tio

n e

rro

r (k

m)

# of packets

Source localization result for an earthquake

16:56:47 Nov 03 2009 @ Mt St Helens

24

Earthquake Source Localization

10

15

20

25

% o

f tr

an

smit

ted

da

ta

0.4

0.6

0.8

Sou

rce

loca

liza

tio

n e

rro

r (k

m)

Packet reception ratio

85%

180 220 260 300 360 220 260 300 3600

5

10

# of packets

% o

f tr

an

smit

ted

da

ta

0

0.2

Sou

rce

loca

liza

tio

n e

rro

r (k

m)

# of packets

Source localization result for an earthquake

16:56:47 Nov 03 2009 @ Mt St Helens

Localization error below 1km, common in volcano seismology

Only 16% data transmission

24