Volcanic Earthquake Timing Using Wireless Sensor Networks GuojinLiu 1,2 RuiTan 2,3 RuoguZhou 2 GuoliangXing 2 Wen-Zhan Song 4 Jonathan M. Lees 5 1 Chongqing University, P.R. China 2 Michigan State University, USA 3 Advanced Digital Science Center, Illinois at Singapore 4 Georgia State University, USA 5 University of North Carolina at Chapel Hill, USA
Volcano Hazards 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] 7% world population live near active volcanoes 20-30 explosive eruptions/year 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 Harvard node 4
Earthquake Timing Node 05 Node 04 Node 06 P-phase 0 1 2 3 Time (second) Earthquake timing Source localization Seismic tomography Key to localization, seismic tomography, etc. Usually done manually, automation is expensive 5
Earthquake Timing Node 09 Node 10? 0 1 2 3 4 Source localization Seismic tomography 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 5
Seismic Signal: Sparsity Original @ 100Hz Time (second) 6
Seismic Signal: Sparsity Original @ 100Hz Sparsity=0.57 Time (second) wavelet K-sparse signal: s s ( k ) s 2 2 < K largest points 5% sparsity = k signal length 6
Seismic Signal: Sparsity Original @ 100Hz Sparsity=0.57 Time (second) Wavelet Sparsity=0.14 Time-frequency domain wavelet K-sparse signal: s s ( k ) s 2 2 < K largest points 5% sparsity = k signal length 6
Seismic Signal: Sparsity Original @ 100Hz Sparsity=0.57 Time (second) Wavelet Sparsity=0.14 Time-frequency domain wavelet K-sparse signal: s s ( k ) s 2 2 < K largest points 5% sparsity = k signal length Observation 1: wavelet sparsifies signal 6
Seismic Signal: Frequency-Time 4-level wavelet transform (length=1600) original thumbnail (length=100) Low-pass band (0, 6.25Hz) P-wave < 5Hz Time (unit: 160ms) 7
Seismic Signal: Frequency-Time 4-level wavelet transform (length=1600) original thumbnail (length=100) Low-pass band (0, 6.25Hz) P-wave < 5Hz Time (unit: 160ms) Original (length=1600) 7
Seismic Signal: Frequency-Time 4-level wavelet transform (length=1600) original thumbnail (length=100) Rough P-phase estimate Time (unit: 160ms) Original (length=1600) 7
Seismic Signal: Frequency-Time 4-level wavelet transform (length=1600) original thumbnail (length=100) Rough P-phase estimate Time (unit: 160ms) Original (length=1600) Observation 2: P-phase estimate from thumbnail 7
Seismic Signal: Diversity Node 1 sparsity=0.1 Node 10 sparsity=0.38 Earthquake 1 8
Seismic Signal: Diversity Node 1 sparsity=0.1 Node 10 sparsity=0.38 Earthquake 1 Node 10 sparsity=0.14 Earthquake 2 8
Seismic Signal: Diversity Node 1 sparsity=0.1 Node 10 sparsity=0.38 Earthquake 1 Node 10 sparsity=0.14 Earthquake 2 Observation 3: sensors have different sparsities 8
Outline Problem statement Approach overview Earthquake timing algorithms Performance evaluation Conclusion 9
Approach Overview Cluster head 10
Approach Overview Cluster head 10
Approach Overview signal sparsity preliminary pick signal sparsity preliminary pick signal sparsity preliminary pick Cluster head signal sparsity preliminary pick signal sparsity preliminary pick Lightweight signal processing algorithms Signal sparsity Preliminary P-phase from thumbnail 10
Approach Overview signal sparsity preliminary pick signal sparsity preliminary pick signal sparsity preliminary pick Cluster head signal sparsity preliminary pick signal sparsity preliminary pick 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 0 1 0 0 0 1 1 0 0 0 1 0 Cluster head 0 1 0 0 0 1 1 0 0 0 1 0 0 0 1 0 1 0 0 1 0 1 0 0 Lightweight signal processing algorithms Signal sparsity Preliminary P-phase from thumbnail Select most informative sensors to TX Compressive sampling & transmission 10
Approach Overview accurate pick Signal reconstruction 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 Sensor selection & compressive sampling Performance evaluation Conclusion 11
Preliminary P-phase Pick 4-level wavelet transform (length=1600) thumbnail (length=100) Time (unit: 160ms) 12
Preliminary P-phase Pick 4-level wavelet transform (length=1600) thumbnail (length=100) preliminary pick Time (unit: 160ms) preliminary pick = 2 4 arg max p thumbnail signal energy after p signal energy before p 12
Preliminary P-phase Pick 4-level wavelet transform (length=1600) thumbnail (length=100) preliminary pick Map thumbnail domain back to original time domain preliminary pick = 2 Time (unit: 160ms) 4 arg max p thumbnail signal energy after p signal energy before p 12
Preliminary P-phase Pick 4-level wavelet transform (length=1600) thumbnail (length=100) preliminary pick Map thumbnail domain back to original time domain preliminary pick = 2 Time (unit: 160ms) 4 arg max p thumbnail signal energy after p signal energy before p Lightweight: O(signal length) Suitable for resource-constrained sensors 12
Outline Problem statement Approach overview Earthquake timing algorithms 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 z z 2 1 V z 3 t t t ray tracing( z, z, V i = i 0 ) 2 1 t 3 z 0 Information-theoretic error metric ( ) ) T 1 E = tr GG scaled Fisher matrix: z i, z 0 14
Impact of Timing on Source Localization Source localization Basis for many volcano monitoring applications Complex non-linear inverse problem sensor position z z 2 1 V z 3 t t t ray tracing( z, z, V i = i 0 ) 2 1 t 3 z 0 Information-theoretic error metric ( ) ) T 1 E = tr GG scaled Fisher matrix: z i, z 0 14
Impact of Timing on Source Localization Source localization Basis for many volcano monitoring applications Complex non-linear inverse problem sensor position source location z z 2 1 V z 3 t t t ray tracing( z, z, V i = i 0 ) 2 1 t 3 z 0 Information-theoretic error metric ( ) ) T 1 E = tr GG scaled Fisher matrix: z i, z 0 14
Impact of Timing on Source Localization Source localization Basis for many volcano monitoring applications Complex non-linear inverse problem sensor position source location z z 2 1 V z 3 t t t ray tracing( z, z, V i = i 0 ) 2 1 t 3 z 0 Information-theoretic error metric volcano model ( ) ) T 1 E = tr GG scaled Fisher matrix: z i, z 0 14
Dynamic Sensor Selection Find a subset of sensors Sto minimize Es.t. i S c i m(sparsity of sensor i) C 15
Dynamic Sensor Selection Find a subset of sensors Sto minimize Es.t. i S c i m(sparsity of sensor i) C unit TX cost 15
Dynamic Sensor Selection Find a subset of sensors Sto minimize Es.t. i S c i m(sparsity of sensor i) C unit TX cost TX volume 15
Dynamic Sensor Selection Find a subset of sensors Sto minimize Es.t. i S c i m(sparsity of sensor i) C unit TX cost TX volume cost budget 15
Dynamic Sensor Selection Find a subset of sensors Sto minimize Es.t. i S c i m(sparsity of sensor i) C unit TX cost TX volume cost budget Brutal-force search 8 seconds on Imote2 for 16 sensors Information gain diminishes for larger clusters 15
Compressive Sampling (CS) n m x n = m random matrix original compressed Apply CS to wavelet coefficients Known TX volume before compression m = 1.5 sparsity n Unselected sensors avoid compression overhead 16
Compressive Sampling (CS) n m x n = m random matrix original compressed Apply CS to wavelet coefficients Known TX volume before compression m = 1.5 sparsity n best trade-off b/w TX volume and signal reconstruction error Unselected sensors avoid compression overhead 16
Outline Problem statement Approach overview Earthquake timing algorithms Performance evaluation Testbedexperiments 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 Execution time (second) 3 2 1 End-to-end delay < 3 seconds 0 1 4 8 12 Sensor ID 18
Error metric Trace-driven Simulation Data traces from 12 sensors on Mt St Helens 30 significant earthquakes in 5.5 months 30 20 10 Lance [SenSys 08] our sensor selection approach 0 5 200 400 600 200 400 600 TX bound (# of pkts) TX bound (# of pkts) # of selected senso ors 11 9 7 Configurable trade-off between system performance and energy consumption 19
Impact of Packet Loss = x Relative reconstruction erro or (%) 30 20 10 0 compressed signal Compressive Sensing Lossy compression: encodes largest wavelet coefficients 65 70 75 80 85 90 95 100 Packet reception ratio (%) reconstructed signal CS is resilient to packet loss! 20
Impact of Packet Loss received = x Relative reconstruction erro or (%) 30 20 10 0 compressed signal Compressive Sensing Lossy compression: encodes largest wavelet coefficients 65 70 75 80 85 90 95 100 Packet reception ratio (%) reconstructed signal CS is resilient to packet loss! 20
Impact of Packet Loss received = x Relative reconstruction erro or (%) 30 20 10 0 compressed signal Compressive Sensing Lossy compression: encodes largest wavelet coefficients 65 70 75 80 85 90 95 100 Packet reception ratio (%) reconstructed signal CS is resilient to packet loss! 20
Accuracy of Timing fine-grained pick on original fine-grained pick on reconstructed 21
Accuracy of Timing fine-grained pick on original fine-grained pick on reconstructed 16% data TX 0.6 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 sensor coordinator cluster 6.7km se ensor / coordinator 0.3 0.2 0.1 STA/LTA detector Bayesian detector 500 nodes on Tungurahua, Ecuador, 2015 [VolcanoSRI project] Sensors Limited capability, large spatial coverage Coordinators Powerful, limited number 0 10 100 200 # earthquakes per day MCU & radio energy ratio TelosB vs. Imote2 23
Earthquake Source Localization % of tra ansmitted data 25 20 15 10 5 0 Source loc calization error (km) 0.8 0.6 0.4 Packet reception ratio 85% 0 180 220 260 300 360 220 260 300 360 # of packets # of packets 0.2 Source localization result for an earthquake 16:56:47 Nov 03 2009 @ Mt St Helens 24
Earthquake Source Localization % of tra ansmitted data 25 20 15 10 5 0 Source loc calization error (km) 0.8 0.6 0.4 Packet reception ratio 85% 0 180 220 260 300 360 220 260 300 360 # of packets # of packets 0.2 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