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
Copyright
c
2016 by SCITEPRESS Science and Technology Publications, Lda. All rights reserved
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
3D
Tomography
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
EN
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
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
t
, ··· , x
t+m1
}
and {x
tm
, ··· , x
t1
} be the samples in the i-th and
(i 1)-th second respectively, then e
i1
=
t1
j=tm
x
j
m
is the average of the (i 1)-th second. The i-th sec-
ond RSAM r
i
is calculated as r
i
=
t+m1
j=t
(x
j
e
i1
)
m
. In
our system, the STA or LTA is continuously updated
based on X
i
=
n1
j=0
r
i j
n
where r
i
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
T
0
. 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
0
50
100
150
200
0 5 10 15 20 25 30
Signal, time (second)
station CGAS, channel BHZ
0
0.5
1
1.5
2
2.5
3
3.5
4
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
}
t
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
i
N (0, σ
2
1
) and the post-change sample x
i
N (0, σ
2
2
).
The logarithm of the likelihood function at time k
can be written as L =
t
i=k+1
h
1
2
ln
σ
2
1
σ
2
2
x
2
i
2
1
σ
2
2
1
σ
2
1
i
.
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
1
, 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.
Figure 13: Waveform of a hammer shock event on sensor
node 09 and SigmaBox 69.
0
2
4
6
8
10
X 10+3
0
1
2
3
4
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
1
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
xo
yo
2
3
4
5
6
7
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
xo
yo
2
3
4
5
6
7
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
38
08
49
11
27
33
24
32
04
17
20
50
22
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