Sensor Network for Real-time In-situ Seismic Tomography

Lei Shi, Wen-Zhan Song, Fan Dong and Goutham Kamath

Georgia State University, Atlanta, Georgia, U.S.A.

Keywords:

Sensor Networks, Design, In-network Processing and Seismic Tomography.

Abstract:

Most existing seismic exploration or volcano monitoring systems employ expensive broadband seismometer

as instrumentation. At present raw seismic data are typically collected at central observatories for post process-

ing. With a high-ﬁdelity sampling, it is virtually impossible to collect raw, real-time data from a large-scale

dense sensor network due to severe limitations of energy and bandwidth at current, battery-powered sensor

nodes. At some most threatening and active volcanoes, only tens of nodes are maintained. With a small

network and post processing mechanism, existing system do not yet have the capability to recover physical

dynamics with sufﬁcient resolution in real-time. This limits our ability to understand earthquake zone or vol-

cano dynamics. To obtain the seismic tomography in real-time and high resolution, a new sensor network

system for real-time in-situ seismic tomography computation is proposed in this paper. The design of the sen-

sor network consists of hardware, sensing and data processing components for automatic arrivaltime picking

and tomography computation. This system design is evaluated both in lab environment for 3D tomography

with real seismic data set and in outdoor ﬁeld test for 2D surface tomography.

1 INTRODUCTION

In present, most existing seismic exploration or vol-

cano monitoring systems use expensive broadband

seismometer and collect raw seismic data for post

processing. Seismic sampling rates for seismic ex-

ploration are usually in the range of 16-24 bit at 50-

200Hz. Collecting all the raw data in real-time from

a large-scale dense sensor network is virtually impos-

sible due to severe limitations of energy and band-

width at current, battery-powered sensor nodes. As

a result, at some most threatening, active volcanoes,

fewer than 20 nodes (Song et al., 2009) are thus main-

tained. With such a small network and post process-

ing mechanism, existing system do not yet have the

capability to recover physical dynamics with sufﬁ-

cient resolution in real-time. This limits our ability

to understand earthquake zone or volcano dynamics

and physical processes under ground or inside vol-

cano conduit systems. Substantial scientiﬁc discov-

eries on the geology and physics of earthquake zone

and active volcanism would be imminent if the seis-

mic tomography inversion could be in real-time and

the resolution could be increased by an order of mag-

nitude or more. This requires a large-scale network

with automatic in-network processing and computa-

tion capability.

To date, the sensor network technology has ma-

tured to the point where it is possible to deploy and

maintain a large-scale network for volcano monitor-

ing and utilize the computing power of each node for

signal processing to avoid raw seismic data collection

and support tomography inversion in real-time. The

methods commonly used today in the procedure of

seismic tomography computation cannot be directly

employed under ﬁeld circumstances proposed here

because they rely on expensive broadband stations

and post processing, also require massive amounts

of raw seismic data collected on a central process-

ing unit. Thus, real-time seismic tomography of high

resolution requires a new mechanism with respect

to low-cost energy efﬁcient system design and in-

network information processing. Then it is possible to

deploy a large-scale network for long-term and com-

pute the tomography in real-time. To clearly address

the challenges in this paper, we ﬁrst give a short de-

scription on the background knowledge of seismic to-

mography based on traveltime principle.

The ﬁrst-arrival traveltime tomography uses P-

wave ﬁrst arrival times at sensor nodes to derive the

internal velocity structure of the subsurface. The ba-

sic workﬂow of traveltime tomography illustrated in

Figure 1 involves four steps: (a) P-wave Arrival Time

Picking. Once an earthquake event happens, the sen-

sor nodes that detect seismic disturbances record the

signals. The P-wave arrival times need to be extracted

118

Shi, L., Song, W-Z., Dong, F. and Kamath, G.

Sensor Network for Real-time In-situ Seismic Tomography.

DOI: 10.5220/0005897501180128

In Proceedings of the International Conference on Internet of Things and Big Data (IoTBD 2016), pages 118-128

ISBN: 978-989-758-183-0

Event Location Ray Tracing Tomography Inversion

Sensor Node

Seismic Rays

Estimated Magma Area

Blocks on Ray Path

Magma

Estimated Event Location

Earthquake Event

(b) (c) (d)

P-wave Arrival Time Picking

P-wave

(a)

Figure 1: Workﬂow of ﬁrst-arrival traveltime tomography.

from the raw seismic data; (b) Event Location. The

P-wave arrival times and locations of sensor nodes

are used to estimate the event hypocenter and origin

time in the volcanic ediﬁce; (c) Ray Tracing. Follow-

ing each event, seismic rays propagate to nodes and

pass through anomalous media. These rays are per-

turbed and thus register anomalous residuals. Given

the source locations of the seismic events and current

velocity model, ray tracing is to ﬁnd the ray paths

from the event hypocenters to the nodes; (d) Tomog-

raphy Inversion. The traced ray paths, in turn, are

used to image a tomography model of the velocity

structure. As shown in Figure 1 the volcano is parti-

tioned into small blocks and the seismic tomography

problem can be formulated as a large and sparse ma-

trix inversion problem.

In traditional seismology, the raw seismic data is

collected for manual analysis including P-wave ar-

rival time picking on the seismograms. Then cen-

tralized methods will process the data and compute

seismic tomography. Keep in mind that our goal is

to design a system which can deliver tomography in

real-time over a large-scale senor network by utilizing

the limited communication ability in the network and

the computation power on the sensor node. To reach

this goal, no raw seismic data should be transmitted

over the network, which requires a light weighted al-

gorithm that can accurately pick the P-wave arrival

time on the sensor nodes locally inside the network.

This paper presents a sensor network system for

real-time in-situ seismic tomography computation.

The design of the sensor network consists of hard-

ware, sensing, data processing, algorithm for auto-

matic arrivaltime picking and so on. This system de-

sign is evaluated both in lab environment for 3D to-

mography with real seismic data set (previous deploy-

ment on San Andreas Fault (SAF) in Parkﬁeld) and in

outdoor ﬁeld test for 2D surface tomography.

The rest of the paper is organized as follows. Sec-

tion 2 shows the system architecture and discusses the

system design in details. In section 3, we evaluate the

signal quality compared with a industry level com-

mercial data logger for seismology and validate the

arrival time picking algorithm with both ﬁeld test and

real seismic data set. Section 4 describes the eval-

uation of the system both in lab environment and in

outdoor ﬁeld test. Section 5 discusses related work.

Section 6 concludes this paper and outlines the future

work.

2 SYSTEM DESIGN

In this section, we give the overview of our system ar-

chitecture and the details of both hardware and soft-

ware design. According to the motivation and require-

ment of the system design discussed above, the spe-

ciﬁc goals of the sensor network system design is as

following:

• Synchronized Sampling. The event location and

travel-time tomography requires the P-wave ar-

rival time of earthquake events. The P-wave ar-

rival time analysis is based on the temporal and

spacial correlation of the recorded signals on sta-

tions. So all stations need to perform synchro-

nized sampling and timestamp the record with

precise UTC time. The synchronization accuracy

should be less than the time interval of sampling

(e.g. 20 millisecond for 50Hz sampling rate).

• Long-term Robust Deployment. To get accurate

event location and high resolution seismic tomog-

raphy, the more data recorded the better result can

be potentially delivered. Since the earthquake ac-

tivities are unpredictable, the long-term robust de-

ployment is necessary to get enough data. Also,

due to the harsh weather conditions for remote de-

ployment, a low-cost energy efﬁcient station with

renewable energy and weatherproof capacity is re-

quired.

• P-wave Arrival Time Picking. As we discussed in

section 1, this sensor network system will send the

arrival time back instead of all raw seismic data.

The system must be able to continuously moni-

tor the signal, detect and pick arrival time in real-

time.

• Online Monitoring and Conﬁguration. To mon-

itor the status of the network, perform the real-

time signal processing and in-situ computation,

the sensor network should be able to respond to

Sensor Network for Real-time In-situ Seismic Tomography

119

external control from the base station for status

report or node conﬁguration. The command and

control needs to be delivered reliably in real-time.

• Distributed Computation Extension. This system

is not limited to be a in-situ signal processing

and data collection framework. In the future, the

system can be used for more complicated seis-

mic analysis that may include cross correlation of

signals between stations, distributed computation

and so on. Those tasks will require more compu-

tation power on each sensor node. An extension

for adding a computation unit is required to make

this system more extensive and general.

2.1 System Architecture

Our system consists of several components. Figure 2

shows the architecture of the sensor network system

design. First, the sensor nodes with seismometers and

RF modules form a mesh network. Each sensor con-

tinuously records the signal, once an event happened,

the sensor will detect it and pick the P-wave arrival

time from the signals. Then the arrival time along

with the station coordinates is delivered to the base

station. The base station is a computer that equipped

with RF module, it runs various tools to process the

received data, compute the tomography, visualize the

result, monitor the network status and conﬁgure the

sensor nodes. This system can deliver either 3D to-

mography through event location, ray tracing and in-

version, or 2D surface tomography with Eikonal to-

mography method (Lin et al., 2009).

Base Station

Event

Location

Ray

Tracing

Tomography

GUI

Eikonal

Tomography

Monitoring &

Configuration

GUI

Arrival times

Status and

Responds

Command and

Request

Sensor nodes each equipped with

seismometer and RF module

Figure 2: Sensor Network System Architecture.

2.2 Hardware Design

Considering the system design goals, our sensor node

design encapsulates all the hardware components in

a weatherproof plastic box. Figure 3 gives the con-

ﬁguration of the sensor node in the ﬁeld. The sensor

node box connected with a 10 Watt solar panel to get

renewable energy. A single-axis 4.5Hz seismometer,

GeoSpace geophone is connected as the sensor com-

ponent. We mounted a 9 dBi omnidirectional antenna

on the box for the communication of a 900MHz RF

module to get a reasonable line-of-sight range. All

the connections are also sealed by weatherproof con-

nectors for the harsh environment. The total weight

of each sensor node station is about 10 pounds which

can be carried by a person for remote deployment.

Figure 3: Sensor node in the ﬁeld.

Figure 4 shows the hardware components inside

the sensor node box, with a dimension of 0.82 ×

0.55 × 0.31 (inch). All our components in the sys-

tem are mounted on a single-layer PCB board. The

core of the system is a TI MSP430F6779 proces-

sor, 25MHz, 512KB of program ROM and 32KB of

SRAM. This processor also provides seven indepen-

dent 24-bit Sigma-Delta ADCs with different inputs

and variable gain.

Figure 4: Hardware components in the box.

The low-power radionova M10478-A2 GPS inter-

face is connected to the processor through UART1 to

provide raw GPS data, and through GPIO 40 to pro-

vide PPS (pulse-per-second) signal capturing. The

GPS interface is used to provide the coordinate of

the sensor node and the timestamps for recorded

data. For wireless communication, we employed the

IoTBD 2016 - International Conference on Internet of Things and Big Data

120

XBee-PRO 900HP module to provide a low-power,

low maintenance, long outdoor range and self orga-

nized wireless network. XBee module takes advan-

tage of the DigiMesh networking protocol. It can run

on dense network operation and support for sleeping

routers for energy efﬁcient. Besides, various point-

to-multipoint conﬁgurations are available for the net-

work. The MSP430 is connected to XBee using

UART2 with 9600 baud that provides 960 Kbps data

rate.

MSP430F6779

UART1 GPIO40 UART2

GPS

TX RX

PPS

XBee

TX RX

RX TX

External antenna

Sensor

CH3 CH2 CH1

ADC0 ADC1 ADC2

RX TX

micro SD card BeagleBone Black

SPI0 GPIO68 SPI1

GPIO72

GPIO70

SOMI

SIMO

SCLK

SPISC

SOMI

SIMO

SCLK

SPISC

Figure 5: The main hardware components connection.

Since the sensor network is designed to sense the

signal, pick and send the P-wave arrival timestamp

back without transmitting the raw seismic data. All

the raw data is stored in a micro SD card for other

post analysis required by seismologists. We use the

DM3D-SF connector and connect the processor with

memory card through SPI0 for SPI communication

and clock, and through GPIO 68 for SPI card select

pin. The node sensor connector is designed to con-

nect up to three channels of seismometer. The node

can connect either to a single-axis or a tri-axis geo-

phone. Both geophones are passive instruments and

the ground motion can generate voltage which is dig-

itized by the ADC module in MSP430.

For the distributed computation extension require-

ment, one BeagleBone Black (BBB) module is con-

nected with the expansion connector to the board

through its SPI0 interface. We use the SPI1 on

MSP430 for SPI communication and clock, the GPIO

72 for SPI card select pin and the GPIO 70 as the

power switch for BBB. The main hardware compo-

nents connection relationships are shown in Figure 5.

2.3 Sensing and Data Processing

Aim to achieve the system design goals, based on our

hardware design, we give the description of the soft-

ware design for sensing and data processing on the

sensor node in this section. Figure 6 illustrates the

framework of the sensing and data processing com-

ponents on the sensor node.

Sensing

Memory Management

SD Card

GPS

Parse

Event

Detection

Timer

GPS Valid

Raw Data

Sensor

XBee

RTC

GPS

Read Write

Buffer full

Picking

Buffer

Process

PPS

Arrvial

times

Request

Responds

Figure 6: Sensing and Data Processing Framework.

Since the accurate event location and high-

resolution tomography are depend on precise timing

by utilizing the temporal and spacial correlation of

recorded signals across stations. The ﬁrst goal of our

system design is synchronized sampling and precise

UTC time timestamp for the recorded data. Our col-

laborators from seismology requires 100Hz sampling

rate on our open nodes. Notice that, the synchronized

sampling is based on time synchronization but not the

same. Synchronized sampling does not only means

that all sensor nodes in the network has the same sam-

ple interval but also sample at the same time point.

In the hardware design, each sensor node employs

a low-cost GPS receiver that provides UTC time in-

formation and PPS signal. The GPS system time from

GPS signal has an accuracy within 50ns referenced

to UTC time, it can be used for time synchroniza-

tion. The problem is that decoding and processing

of the GPS message can generate delays and degrade

the synchronization accuracy. Instead, we use PPS

signals to synchronize the RTC. To achieve the syn-

chronized sampling, we designed a Timer component

to maintain RTC with millisecond resolution. Then

when the system catches the ﬁrst valid PPS interrupt,

the timer is reset and keep counting on milliseconds.

When next valid PPS interrupt is captured, if the timer

is not in exact thousand milliseconds, the timer will be

reset and the RTC will be synchronized properly.

Notice that, the GPS signal can disappear or the

PPS signal can not ﬁre properly. If the time period

without GPS signal or valid PPS signals is long, the

sampling across sensor nodes might not be synchro-

Sensor Network for Real-time In-situ Seismic Tomography

121

nized. The sensor node tags every second of data with

a timestamp and a ﬂag. The timestamp represents the

time point corresponding to the ﬁrst sample in this

second. The ﬂag indicates that whether the samples

in this second is under valid time synchronization or

not. The system will tag the second of data invalid

synchronization if: (1) there is no GPS signal for 60

seconds; (2) there is no valid PPS interrupts for 20

seconds.

The Sensing component samples the sensor with

10 milliseconds sample interval according to the

Timer component. There is a small circular buffer

to sample one hundred samples (one second data un-

der 100Hz) from sensor. Once this buffer is full.

The Sensing component will send it to memory man-

agement to write the buffer into micro SD card with

proper timestamp and ﬂag. Also, the Event Detection

component takes this buffer and perform the event de-

tection processing, if one event is detected, the related

buffered data is processed by Picking component to

get the arrival time and send it through XBee mod-

ule. Details about the event detection and arrival time

picking are discussed in section 2.4. There is another

module Process that processes the requests from base

station and send responds back for status monitoring

and network conﬁguration. More details about this

can be found in section 2.5.

2.4 P-wave Arrival Time Picking

Primary waves (P-waves) are the seismic waves that

travel faster than any other waves through the earth.

P-waves arrive at the seismic sensors ﬁrst and the ar-

rival time of P-waves are essential to the ﬁrst-arrival

traveltime tomography. Figure 7 shows the seismo-

grams from two seismometers deployed in Parkﬁeld

when an event happens. The vertical lines represent

manual pickings of the P-wave arrival times. Due to

the different wave propagation delays, the P-wave ar-

rival times on sensors are different. In local seismic

tomography, the scale of the ﬁeld is up to tens of kilo-

meters and the maximum difference of the P-wave ar-

rival times among sensors is about several seconds,

so that the accuracy of the picking is signiﬁcant. Be-

sides, manual analysis of seismograms and picking of

arrival times require post processing of the data and

are very time consuming, especially in a large sensor

network. To avoid raw seismic data transmission and

meet the real-time requirements, the system demands

an on-line automatic event detection and P-wave ar-

rival time picking method that runs on each sensor

node.

In this paper, we proposed a two-step method for

P-wave arrival time picking. (1) Event Detection

station GOBI, channel BHZ

0 1 2 3 4 5 6

Time (second), 12:18:46 to 12:18:52 Oct 4, 2001

station GULY, channel BHZ

Figure 7: The seismogram from BHZ channel of two seis-

mometers in Parkﬁeld when an event happens. The vertical

lines indicate the manual pickings of P-wave arrival times.

which continuously scanning the samplings from the

sensor with a sliding window, claims if there is an

event (change point) happens and extract a segment

of signals around the change point; (2) Arrival Time

Picking which takes the segment of signals from step

(1), picks the exact change point (arrival time) from it

and send the P-wave arrival time to coordinator node.

Figure 8 illustrates how the proposed method works.

(1) Event Detection

Sensor Node

Sliding Window

(2) Arrival Time Picking

event

detected

accurate P-wave

arrival time

Samplings

Figure 8: Two step P-wave arrival time picking.

In the ﬁrst step, the goal is to continuously check

weather there is a change point in the signal stream

that is probably an event. The STA/LTA (short-term

average over long-term average) algorithm (Murray

and Endo, 1992; Song et al., 2009) is employed for

event detection here because it is fast to monitor the

signal and roughly ﬁnd where an event happens. To

describe STA/LTA algorithm, we need to ﬁrst intro-

duce the concept of RSAM (Realtime Seismic Am-

plitude Measurement), which is widely used in seis-

mology. The RSAM is calculated on raw seismic data

samples per second. Assume the sampling rate of the

signal is m (samples per second), let {x

, ··· , x

t+m−1

}

and {x

t−m

, ··· , x

t−1

} be the samples in the i-th and

(i − 1)-th second respectively, then e

i−1

∑

t−1

j=t−m

is the average of the (i − 1)-th second. The i-th sec-

ond RSAM r

is calculated as r

∑

t+m−1

j=t

−e

i−1

)

. In

our system, the STA or LTA is continuously updated

based on X

∑

n−1

j=0

i− j

where r

is i-th second RSAM;

n is the STA or LTA time window size (in seconds).

The ratio of STA over LTA is continuously moni-

IoTBD 2016 - International Conference on Internet of Things and Big Data

122

tored. Once the ratio exceeds the threshold, an event

is detected. A sliding LTA window with a STA win-

dow keep moving second by second and calculating

the STA/LTA ratio. If the threshold b is reached, a

change point is detected at T , the signal in the win-

dow of [T −a, T +b] are extracted and the arrival time

picking algorithm will pick the accurate arrival time

from it. Then the sliding window continue moving

from T and calculating the STA/LTA ratio, since the

event usually lasts for a period of time, the STA/LTA

ratio will be over the threshold for a while until time

. In our implementation, the STA and LTA window

are 1 and 4 seconds respectively; the signal window

for picking is 3 seconds where a = 1 and b = 2. These

parameters are all conﬁgurable in the network with

base station. This setting with threshold 2 can per-

form event detection very well for picking as shown

in section 3.2. Figure 9 gives an example of event de-

tection result and the detection length of a real earth-

quake event.

-200

-150

-100

-50

100

150

200

0 5 10 15 20 25 30

Signal, time (second)

station CGAS, channel BHZ

0.5

1.5

2.5

3.5

0 5 10 15 20 25 30

STA/LTA ratio, time (second)

station CGAS, channel BHZ

T T’

Figure 9: Event detection example on a real earthquake

event.

From the seismograms in Figure 7, one can see

that there is a big difference on the amplitude vari-

ance of the signal before and after the arrival of P-

waves, the P-wave arrival time is a change point of

the variance of the signal amplitude. Based on this

observation, a P-wave picking method in step two

is proposed in this paper by utilizing the maximum-

likelihood (ML) estimation to estimate the variance

of the signal amplitude following a statistical model.

Without loss of generality, we assume that both

the pre- and post-change (before and after P-wave ar-

rival) signals follow a normal distribution but with

different variances. Let {x

}

i=1

be the continuous se-

quence of samples from 1 to t, then the pre- and post-

change sample has a normal distribution with zero

mean respectively. Then the pre-change sample x

∼

N (0, σ

) and the post-change sample x

∼ N (0, σ

The logarithm of the likelihood function at time k

can be written as L =

∑

i=k+1

−



−



The arrival time picking is to ﬁnd the exact change

point k

∗

in [T −a, T +b] which maximize the function

value of L. Due to the length limit, we omit the full

mathematical derivation process and equations for the

P-wave picking calculation.

2.5 Online Monitoring and

Conﬁguration

Considering the complexity and remoteness of en-

vironment monitoring, online status monitoring and

sensor node conﬁguration are highly desired. With

online monitoring, users can easily get the status of

the sensor nodes in the network. This is very helpful

when the deployment is initiating, one can monitor all

the sensor nodes in the network remotely without ac-

tually visiting them remotely. There are two modes

in the system, test and deploy. When the deployment

starts, the node will start with test mode, and it will re-

port the status periodically. After a while, if the node

status is normal, users can switch the nodes to deploy

mode where the sensor node only report the status if

requested. The status report consists of the GPS status

(satellite numbers, latitude, longitude, altitude), the

sensing status (number of events detected, number of

seconds recorded), the power status (solar panel input

voltage and battery voltage). Besides, many parame-

ters need to be conﬁgurable in the sensor node. For

example, the window size and threshold in the event

detection and picking algorithms. Since different kind

of events have different properties, different parame-

ters could identify various classes of events according

the interesting of seismologists. Also, the sensing pa-

rameters such as channel, data resolution, sensor sta-

tus and reference voltage gain can also be conﬁgured.

In our system design, the sensor node only sends

the arrival time with station information back by de-

fault. One problem is that in different ﬁeld or with dif-

ferent interests for events, the parameters for event de-

tection and picking can be varied. After deployment,

users need to know whether the arrival time picking

is accurate or not. Thus, a stream option is added into

the system. When the stream option is on, the sen-

sor node will send the raw stream data in the picking

buffer with the arrival time. Users can visualize it on

base station to check the picking accuracy on base sta-

tion in real-time. Figure 10 shows the stream data and

arrival time picking from the monitoring and conﬁgu-

ration tool on the base station. Besides, seismologists

might be interesting in the raw data for other analysis

in the deployment period. Then can not afford to visit

the sensor node remotely all the time. Another fea-

ture in this system is that users can download the data

Sensor Network for Real-time In-situ Seismic Tomography

123

from a any node by specifying the start and end time

point.

Figure 10: Stream data with arrival time picking.

3 DATA QUALITY AND PICKING

ACCURACY

Before the ﬁeld deployment and end-to-end tomog-

raphy computation test of the system, we conducted

several tests to verify the quality of recorded data with

the sensor node and the accuracy of the arrival time

picking mechanism.

3.1 Data Quality

The scientiﬁc value of the data is the ﬁnal and most

important measurement of the sensor network sys-

tem. The ﬁrst test here is to see weather this system

can provide scientiﬁcally meaningful data to seismol-

ogists. Since it is not easy to ﬁnd a place to record

earthquake events and it might take long time to val-

idate. With the suggestion from seismologists, we

conduct a hammer shock test that is commonly used

by the experts for preliminary. This test is to use a

hammer to hit on the ground to generate seismic wave

propagation. The signal from a hammer shock is not

so different from an earthquake except the energy is

smaller.

Figure 11: SigmaBox conﬁguration.

To validate the data quality, the test is conducted

to compare the data recorded by our sensor node and

a current state of the art commercial seismic acquisi-

tion system called SigmaBox. The SigmaBox is de-

signed by iSeis Corporation

, shown in Figure 11. In

the test, 7 sensor nodes and 4 SigmaBox are deployed.

Four sensor nodes are placed with SigmaBoxes side

by side to compare the recorded data quality, see Fig-

ure 12. The distance between each pair of nodes is 10

meters. We used the hammer to hit the ground near

the SigmaBox 70 and sensor node 18 for 20 times. In

Figure 13, we can see the recorded data for a hammer

shock event by our sensor node and SigmaBox. The

SNR is similar between two data record and the seis-

mologist were satisﬁed with the data quality overall.

Figure 12: SigmaBox and sensor node deployment.

-2

X 10+5

-2

-1

X 10-3

1350.0 1350.2 1350.4 1350.6 1350.8 1351.0 1351.2 1351.4 1351.6 1351.8 1352.0

Node 09

Node 69

Figure 13: Waveform of a hammer shock event on sensor

node 09 and SigmaBox 69.

X 10+3

X 10-5

0 10 20 30 40 50

Node 09

Node 69

Figure 14: Spectrum of the hammer shock event on sensor

node 09 and SigmaBox 69.

The spectrum of the waveform in Figure 13 is

http://www.iseis.com

IoTBD 2016 - International Conference on Internet of Things and Big Data

124

shown in Figure 14. We can see that the spectrum dis-

tribution is similar between two signals. This further

validate the data quality of our sensor node. Notice

that the SigmaBox costs about $3K, while the sensor

node costs less than $1K.

3.2 P-wave Arrival Time Picking

Accuracy

The example in Figure 13 shows the arrival time pick-

ing of our algorithm on sensor node and SigmaBox

sensed data. Notice that the sampling rate of

SigmaBox was 500Hz in the test. In the example, the

time difference between two pickings is 6 millisec-

ond, which is smaller than the sampling interval of

our sensor node. The average time difference of all

pair of pickings in this test is 4.2 millisecond. This

test shows that the algorithm can deliver similar ar-

rival time result on different node. But it only means

that the recorded data quality from two kinds of nodes

is similar to perform detection and picking algorithm.

To validate the accuracy of the picking algorithm, the

algorithm should perform on the real data set with

manual pickings from experts as the reference.

Figure 15: Audio to sensor channel adapter.

We used a real data set obtained from seismolo-

gists. This data set was recorded from a previous de-

ployment on San Andreas Fault (SAF) at Parkﬁeld.

The deployment is from Jan 1, 2000 to Dec 31, 2002

with 61 stations. The data set has been cut into short

waveforms that contain events with the manual pick-

ings. Then the problem is that how can we send the

waveforms to sensor node for validation. We made

an adapter from audio input to channel 0 as shown

in Figure 15. Then waveforms were converted into

audio wave and can be sent to the sensor node by any

audio player on a computer, cellphone or tablet. There

are totally 4478 arrival times picked by the algorithm

from the data set. About 91% picking errors of our

algorithm are within 0.2 seconds. The mean value

and the standard deviation of the difference between

our pickings and manual pickings are 0.043 and 0.23.

This is comparable with some recent method in seis-

mology literature (Zhang et al., 2003).

4 SYSTEM EVALUATION

In this section, we conduct two experiments to eval-

uate the sensor network system for both 3D and 2D

surface tomography.

4.1 Parkﬁeld 3D Tomography

From the discussion of previous section, we use an

adapter to send the waveform from computer to our

sensor nodes simulating the sampling process. The

sensor node then process the data, picked the arrival

times and send to a base station set in the lab. After

the base station received some arrival times, it should

compute the event location. Notice that this computa-

tion is an online process, because in real deployment,

one can not predict when and how many arrival times

can be received since the earthquake activity is not

predictable.

8 10 12 14 16 18 20

20 22 24 26 28 30

●

●●

●

8 9 10 11 12 13 14 15 16 17 18 19 20

20 21 22 23 24 25 26 27 28 29 30

Model: vp1

Layer 12 of 72 layers along Z

Resolution: 120x160

(a) Vp at depth = 2km

8 10 12 14 16 18 20

20 22 24 26 28 30

●

●●

●

8 9 10 11 12 13 14 15 16 17 18 19 20

20 21 22 23 24 25 26 27 28 29 30

Model: vp1

Layer 20 of 72 layers along Z

Resolution: 120x160

(b) Vp at depth = 4km

Figure 16: Horizontal slices of the P-wave velocity at

depths of 2, 4 km. The fault is located around X=13.5km.

The base station only receives the arrival times

from the sensor nodes and has the knowledge that

which picking is from which node. In all of these

pickings, there might be false alarms or some small

and remote event is only detected by few sensor

nodes. As we known, to estimate an event location

and origin time, at least four pickings from different

sensor nodes are required, the event detected only by

one or two nodes is impossible to be located. Also,

more pickings from different sensor nodes for one

event usually lead to a better estimation. Thus there

are two steps in event location, (1) Event Identiﬁca-

tion where the base station identiﬁes how many events

existing in a series of arrival times received and which

pickings belong to the same event; (2) Location Esti-

mation which uses Geiger’s method to estimate the

event location from the arrival times of that event.

Sensor Network for Real-time In-situ Seismic Tomography

125

After identiﬁed the events and computed the event

locations, base station will do ray tracing based on the

event information and the station coordinates received

with the arrival times, followed by the 3D tomography

inversion.

Figure 16 shows the tomography result from the

base station. It is easy to see that the velocity model

is different on different side of SAF. The dots in the

tomography indicate the event locations estimated on

base station. A scientiﬁc fact is that the events of-

ten happen around the fault which is veriﬁed by our

result. We can see that the fault feature is easy to get

from the tomography result. This result is comparable

with the previous research on the Parkﬁeld tomogra-

phy (Zhang et al., 2009).

4.2 Hammer Shock Field Test

To verify the sensor network system in outdoor ﬁeld,

we conduct a ﬁeld test and created the event with

hammer shock on the ground to generate the surface

wave. In this case, we can control the location of the

event source and it is easy to verify if the recorded

data is meaningful, the arrival time pickings is correct

and the 2D surface tomography result is validated.

Figure 17: Hammer shock test deployment.

In the previous discussion, the hammer shock test

has already been used for data quality and arrival time

picking validation. From that test, we found that

on the soil ground, the hammer shock can generate

waves propagated up to around 30 meters, depends on

how hard the hammer hit the ground. Thus, 25 sen-

sor nodes are deployed on a 20×20 meter area with 5

meter space between the adjacent sensor nodes. Fig-

ure 17 shows the deployment of 25 sensor nodes. In

the area we deployed, the upper half of it covered by

wet soil under the tree while the other half is covered

by drier soil under the sunshine in day time. The rea-

son we choose this area is that we would like to see

the difference from Eikonal tomography based on the

property of the different acoustic wave propagation

speed in wet and dry soil.

After the deployment done, the base station mon-

itored the status of all sensor nodes and told us when

all nodes stared working normally. Then we created

38 49 27 33 24

19 08 11 48 32

07 36 15 17 04

13 45 29 50 20

26 42 05 16 22

5 m

Figure 18: Deployment map of sensor nodes.

5 hammer shocks (events) beside each station, totally

125 events were created. The deployment map of the

sensor nodes is shown in Figure 18.

Finally, after all hammer shocks done, the base

station received more than 2000 arrival times and

computed the 2D surface tomography with Eikonal

tomography method. Before showing the tomogra-

phy result, we take a closer look at one seismic event

recorded by the sensor network and the arrival times

picked out of this event.

62.50 62.55 62.60 62.65 62.70 62.75 62.80 62.85 62.90 62.95 63.00 63.05 63.10 63.15 63.20 63.25 63.30 63.35 63.40 63.45 63.50

Figure 19: Hammer shock event captured.

The hammer shock event captured in Figure 19

was generated by hit beside node 24, which located on

the upper right corner in the deployment map. From

the recorded signal and picked arrival times shown in

Figure 19, node 24 got the earliest arrival time and

the further nodes got the relatively delayed arrivals,

which shows the wave propagation in the deployed

area. Within such a small area, this further veriﬁed the

synchronized sampling accuracy of our sensor net-

work system.

In this test, the base station received totally 2012

arrival times picked on the sensor nodes. Some events

IoTBD 2016 - International Conference on Internet of Things and Big Data

126

are not picked on some nodes. The reason is that the

sensor node can not get good event signal, it depends

on how hard the hammer hit the ground and how far

the sensor node is from the event location.

km/s

Figure 20: 2D surface wave tomography.

Out of 2012 arrival times, the base station iden-

tiﬁed 96 events with 1905 arrival times. Figure 20

shows the 2D surface wave tomography delivered by

the base station. According to the research on acous-

tic wave propagation in soil (Oelze et al., 2002), the

impedance mismatch from the water to air is much

greater than the water to soil frame. Thus, more satu-

rated the soil is, slower the acoustic wave propagates

in it. As we discussed above, the upper side of the

deployed area contains more water in the soil. This

observation is shown in the ﬁnal tomography result.

5 RELATED WORK

Static tomography inversion for 3D structure, applied

to volcanoes and oil ﬁeld explorations, has been ex-

plored since the late 1970’s (Iyer and Dawson, 1993;

Vesnaver et al., 2003; Lees, 2007). In volcano ap-

plications, tomography inversion used passive seis-

mic data from networks consisting of tens of nodes,

at most. The development and application to volca-

noes include Mount St. Helens (Lees, 1992; Lees

and Crosson, 1989; Waite and Moranb, 2009), Mt.

Rainier (Moran et al., 1999), Kliuchevskoi, Kam-

chatka, Russia (Lees et al., 2007), and Unzen Vol-

cano, Japan (Ohmi and Lees, 1995). At the Coso

geothermal ﬁeld, California, researchers have made

signiﬁcant contributions to seismic imaging by coor-

dinating tomography inversions of velocity (Wu and

Lees, 1999), anisotropy (Lees and Wu, 1999), atten-

uation (Wu and Lees, 1996) and porosity (Lees and

Wu, 2000).

Sensor network has been deployed for monitoring

in many different areas. In (Cerpa et al., 2001), the

sensor network was deployed to collect dense envi-

ronmental and ecological data about populations of

rare species and their habitats. Another sensor sys-

tem was used by the researchers to monitor the habi-

tat of the Leach’s Storm Petrel at Great Duck Is-

land (Szewczyk et al., 2004). The Zebranet (Juang

et al., 2002) project uses sensor network nodes at-

tached to zebras to monitortheir movements via GPS.

It is composed of multiple mobile nodes and a base

station with occasional radio contact. The sensor net-

work was also used to monitor the bridge health (Kim

et al., 2006; Chebrolu et al., 2008) and weather con-

dition as well (Hartung et al., 2006).

For volcano monitoring, The ﬁrst volcanic mon-

itoring work using WSN was developed in July

2004 (Werner-Allen et al., 2006), by a group of

researchers from the Universities of Harvard, New

Hampshire, North Carolina, and the Geophysical In-

stitute of the National Polytechnic School at Reventa-

dor in Ecuador. Data collection was performed with

continuous monitoring during 19 days.

In 2008, a smart solution was proposed for collect-

ing reliable information aiming to improve the collec-

tion of real-time information. The sensor network was

deployed on Mount St. Helens (Song et al., 2009) for

volcano hazard monitoring and run for months.

6 CONCLUSION

In this paper we presented a sensor network system

that performs in-situ signal processing and obtained

3D or 2D surface tomography in real-time. The hard-

ware and software design of the system focused on

delivering a low-cost, energy efﬁcient and reliable

system to monitor and image the earthquake zone or

active volcano. Several tests and experiments were

conducted to show: (1) the recorded data quality is

similar to current commercial industry level product;

(2) the system can deliver the validated tomography

result. This sensor network system marks the collabo-

ration between geophysicists and computer scientists

that provided opportunities to introduce new technol-

ogy for geophysical monitoring. The design and pre-

sented here has broader implication beyond tomogra-

phy inversion and can be extended to oil and natural

gas exploration.

Our future plan is to deploy a larger scale sensor

network on a real volcano for a long-term run, and

further verify the correctness, the efﬁciency and the

robustness of the system. Another plan in the future

involves the development of distributed algorithm to

Sensor Network for Real-time In-situ Seismic Tomography

127

compute the tomography in a fully distributed manner

without a base station by utilizing the extension of the

computation unit in the sensor node.

REFERENCES

Cerpa, A., Elson, J., Estrin, D., Girod, L., Hamilton, M.,

and Zhao, J. (2001). Habitat Monitoring: Application

Driver for Wireless Communications Technology. In

1st ACM SIGCOMM Workshop on data communica-

tion in Latin America and the Caribbean.

Chebrolu, K., Raman, B., Mishra, N., Valiveti, P. K., and

Kumar, R. (2008). BriMon: A Sensor Network Sys-

tem for Railway Bridge Monitoring. In The 6th An-

nual International Conference on Mobile Systems, Ap-

plications and Services (MobiSys).

Hartung, C., Han, R., Seielstad, C., and Holbrook, S.

(2006). FireWxNet: A Multi-Tiered Portable Wireless

System for Monitoring Weather Conditions in Wild-

land Fire Environments. In The 4th International Con-

ference on Mobile Systems, Applications, and Services

(MobiSys 2006).

Iyer, H. M. and Dawson, P. B. (1993). Imaging volcanoes

using teleseismic tomography. Chapman and Hall.

Juang, P., Oki, H., Wang, Y., Martonosi, M., Peh, L., and

Rubenstein, D. (2002). Energy Efﬁcient Computing

for Wildlife Tracking: Design Tradeoffs and Early

Experiences with ZebraNet. Proc. 10th international

conference on Architectural support for programming

languages and operating systems.

Kim, S., Pakzad, S., Culler, D., Demmel, J., Fenves, G.,

Glaser, S., and Turon, M. (2006). Wireless sensor net-

works for structural health monitoring. In Proc. 4th

ACM conference on Embedded networked sensor sys-

tems (SenSys).

Lees, J. M. (1992). The magma system of Mount St. He-

lens: non-linear high-resolution P-wave tomography.

Journal of Volcanology and Geothermal Research,

53:103–116.

Lees, J. M. (2007). Seismic tomography of magmatic sys-

tems. Journal of Volcanology and Geothermal Re-

search, 167(1-4):37–56.

Lees, J. M. and Crosson, R. S. (1989). Tomographic In-

version for Three-Dimensional Velocity Structure at

Mount St. Helens Using Earthquake Data. Journal of

Geophysical Research, 94(B5):5716–5728.

Lees, J. M., Symons, N., Chubarova, O., Gorelchik, V.,

and Ozerov, A. (2007). Tomographic images of kli-

uchevskoi volcano p-wave velocity. AGU Monograph,

172:293–302.

Lees, J. M. and Wu, H. (1999). P wave anisotropy,

stress, and crack distribution at Coso geothermal

ﬁeld, California. Journal of Geophysical Research,

104(B8):17955–17973.

Lees, J. M. and Wu, H. (2000). Poisson’s ratio and poros-

ity at Coso geothermal area, California. Journal of

Volcanology and Geothermal Research, 95(1-4):157–

173.

Lin, F.-C., Ritzwoller, M. H., and Snieder, R. (2009).

Eikonal tomography: surface wave tomography by

phase front tracking across a regional broad-band

seismic array. Geophysical Journal International,

177(3):1091–1110.

Moran, S. C., Lees, J. M., and Malone, S. D. (1999). P wave

crustal velocity structure in the greater Mount Rainier

area from local earthquake tomography. Journal of

Geophysical Research, 104(B5):10775–10786.

Murray, T. L. and Endo, E. T. (1992). A real-time seismic-

amplitude measurement system (rsam). volume 1966

of USGS Bulletin, pages 5–10.

Oelze, M. L., O’Brien, W. D., and Darmody, R. G. (2002).

Measurement of attenuation and speed of sound in

soils. Soil Science Society of America Journal,

66(3):788–796.

Ohmi, S. and Lees, J. M. (1995). Three-dimensional P- and

S-wave velocity structure below Unzen volcano. Jour-

nal of Volcanology and Geothermal Research, 65(1-

2):1–26.

Song, W.-Z., Huang, R., Xu, M., Ma, A., Shirazi, B., and

Lahusen, R. (2009). Air-dropped Sensor Network for

Real-time High-ﬁdelity Volcano Monitoring. In The

7th Annual International Conference on Mobile Sys-

tems, Applications and Services (MobiSys).

Szewczyk, R., Polastre, J., Mainwaring, A., Anderson, J.,

and Culler, D. (2004). Analysis of a Large Scale

Habitat Monitoring Application. In Proc. 2nd ACM

Conference on Embedded Networked Sensor Systems

(SenSys).

Vesnaver, A. L., Accaino, F., Bohm, G., Madrussani, G.,

Pajchel, J., Rossi, G., and Moro, G. D. (2003). Time-

lapse tomography. Geophysics, 68(3):815–823.

Waite, G. P. and Moranb, S. C. (2009). VP Structure of

Mount St. Helens, Washington, USA, imaged with lo-

cal earthquake tomography. Journal of Volcanology

and Geothermal Research, 182(1-2):113–122.

Werner-Allen, G., Lorincz, K., Johnson, J., Lees, J., and

Welsh, M. (2006). Fidelity and Yield in a Volcano

Monitoring Sensor Network. In Proc. 7th USENIX

Symposium on Operating Systems Design and Imple-

mentation (OSDI).

Wu, H. and Lees, J. M. (1996). Attenuation structure of

Coso geothermal area, California, from wave pulse

widths. Bulletin of the Seismological Society of Amer-

ica, 86(5):1574–1590.

Wu, H. and Lees, J. M. (1999). Three-dimensional P and

S wave velocity structures of the Coso Geothermal

Area, California, from microseismic travel time data.

Journal of Geophysical Research, 104(B6):13217–

13233.

Zhang, H., Thurber, C., and Bedrosian, P. (2009). Joint

inversion for Vp, Vs, and Vp/Vs at SAFOD, Park-

ﬁeld, California. Geochem. Geophys. Geosyst.,

10(11):Q11002+.

Zhang, H., Thurber, C., and Rowe, C. (2003). Auto-

matic P-Wave Arrival Detection and Picking with

Multiscale Wavelet Analysis for Single-Component

Recordings. Bulletin of the Seismological Society of

America, 93(5):1904–1912.

IoTBD 2016 - International Conference on Internet of Things and Big Data

128