AN INVERSE SENSOR MODEL FOR EARTHQUAKE DETECTION

USING MOBILE DEVICES

Thomas Collins and John P. T. Moore

Thames Valley University, London, U.K.

Keywords:

Environmental monitoring and control, Earthquake detection, Nonlinear signals and systems.

Abstract:

We describe a sensory framework to be used for the purposes of earthquake detection using minimal cost,

accelerometer equipped, hardware units. Combining techniques from mobile robotics this model is intended to

address the current issue in the ﬁeld whereby high ﬁdelity hardware units tuned to detect speciﬁc characteristics

such as wave features and/or high ﬁdelity event models derived from data analysis are required for such

detection. In this paper we present and contextualise the architecture under construction in addition tooutlining

the salient elements of the problem we are addressing.

1 INTRODUCTION

At the onset of an earthquake, dedicated detection

systems are capable of issuing alerts thereby provid-

ing valuable time for people further from the event

epicentre to take action to protect themselves. This is

possible as the means employed to facilitate the trans-

mission of alert information is generally faster than

the speed of seismic waves. However this capability

obviously relies on the existence of a monitoring net-

work. Within the ﬁeld of earthquake engineering two

broad approaches have emerged toward the creation

of such networks. One approach consists of creating

banks of accurate units capable of detecting seismic

characteristics because they have been tuned to detect

speciﬁc characteristics such as ﬂuctuating ambient

seismic noise level. Countries such as Japan have suc-

cessfully developed early warning systems based on

such techniques. Unfortunately it is not always pos-

sible to deploy such technology in earthquake prone

areas around the world for varying degrees of polit-

ical, technical and ﬁnancial considerations. Where

such situations manifest themselves the use of com-

modity hardware as the foundation of a seismic de-

tection network is a viable alternative. This high-

lights the second domain approach which consists of

employing the distributed computing paradigm where

the individual physical nodes are typically Laptop or

Desktop computers equipped with sensors such as ac-

celerators. For example the Network for Earthquake

Engineering Simulation Cyberinfrastructure Centre

(NEESit) has utilised the accelerometers in Apple

Macintosh laptops to develop an educational and re-

search platform for measurement and recording of vi-

brations and dynamic responses. Likewise the Quake-

Catcher Network (QCN)2 links existing laptop and

desktop computers with the aim of forming a large

earthquake monitoring system.

While the availability of Laptop/Desktop based

systems does remove a number of obstacles barring

the realisation of detection systems the installations

in themselves suffer from a number of manifest prob-

lems. For example the existence of networking infras-

tructure capable of linking the individual machines,

the potential availability of technically skilled people

(or appropriate training programmes) to operate the

machines, the ability to quickly disseminate alerts is

a required component of any such system.

Our aim is to directly address this problem by

supplementing traditional Laptop/Desktop based ap-

proaches using mobile phone technology. The core

design goal is to develop a portable system which is

capable of running on constrained and resource lim-

ited hardware thereby allowing earthquake detection

in sparsely seismically-instrumented regions. There-

fore a key facet of the system is the development of a

signal processing model which does not have the dis-

tributed quantitative analysis requirementsof the Lap-

top/Desktop techniques mentioned above and is capa-

ble of operation in the context of low entropy sensory

information. In this paper we present such a model,

illustrating its core features and operational charac-

teristics, and presenting initial results illustrating the

competencies of the model and ﬁnally highlight areas

Collins T. and P. T. Moore J. (2010).

AN INVERSE SENSOR MODEL FOR EARTHQUAKE DETECTION USING MOBILE DEVICES.

In Proceedings of the 7th International Conference on Informatics in Control, Automation and Robotics, pages 81-86

DOI: 10.5220/0002911800810086

 SciTePress

of future work.

2 SEISMIC EVENT SIGNAL

DETECTION/HANDLING

There are two broad approaches to the detection

and/or handling of seismic event signals. One ap-

proach consists of creating dedicated accurate sen-

sory units tuned to detect speciﬁc signal characteris-

tics therefore making the units capable of detecting

emergent seismic characteristics e.g. ﬂuctuating am-

bient seismic noise level. The other approach consists

of employing knowledge based reasoning and estab-

lished signal processing techniques to derive signal

processing techniques formed from the feature analy-

sis of historical data.

Hardware based detectors use signal averaging

techniques in an attempt to achieve an optimum sig-

nal to noise ratio which is capable of determining a

true seismic event from a false positive. The ability

of such sensors is directly related to the noise model

that has been pre-determined and incorporated into

the units. Such noise models are generally formed

throughtraditional signal manipulationtechniques i.e.

the statistical analysis of an appropriate domain char-

acteristic function. However to be useful in a practical

sense this noise model must be determined for every

new installation of such the units and limits the detec-

tion threshold thereby reducing the overall effective-

ness of such units (Newmark and Rosenblueth, ).

The direct application of Machine Learning and/or

statistical techniques, typically realised in software, in

the form of knowledge-basedreasoning is an alternate

approach which provides for a level of ﬂexibility in

the detecting of seismic activities. Such systems are

exempliﬁed in (Hewitt, 1992; Zareian and Krawin-

kler, 2009). In this case the detection threshold as-

sociated with the system is not directly dependent on

a physical characteristic such as seismic noise level.

Rather relevant characteristics are determined through

the analysis of domain expertise in the form of histori-

cal data and knowledge acquired from human experts.

The Laptop/Desktop based systems outlined previ-

ously typically employ such techniques. The gathered

information is employed to construct an operational

model which is used to evaluate the sensory infor-

mation received from the Laptop/Desktop sensor(s).

The overall success of such approaches however is

largely dependent on characteristics such as the se-

lection of appropriate domain classiﬁcations and re-

ﬁnement/training of the derived model.

Within the context of the domain we are address-

ing the realisation of an efﬁcient and expressive sig-

nal processing mechanism is paramount to the overall

performance of detection system. Unfortunately nei-

ther of the existing techniques outlined above are di-

rectly usable for what we need to achieve. Techniques

associated with tuned hardware units are not usable

because the hardware units we are concerned with

are standard mobile phone handsets meaning that the

modiﬁcation of same would require speciﬁc technical

expertise and the availability of specialised hardware

which is not a feasible goal for the intended deploy-

ment locations. In addition the application of existing

’knowledge based’ techniques is not directly possible

because of the data requirements both in terms of con-

structing an initial model and subsequent data propa-

gation throughout the network.

From an operational perspective, any signal pro-

cessing technique must consider real time operation

as being paramount. In addition there should be no

requirement for historical knowledge. However any

such information, if available, should be easily incor-

porated into the model developed using the process-

ing technique. Finally the technique must accommo-

date low entropy sensory information.

In evaluating these requirements techniques from

a number of varied domains such as speech recog-

nition e.g. (Vargas et al., 2001), and telecoms e.g.

(Murooka et al., 2001) and mobile robotics e.g.

(Ehlers et al., ) were considered. After domain eval-

uation we determined that the problem that is closest

to the problem we are addressing in developing our

signal handling model is the ﬁeld of Occupancy Grid

based robotic mapping.

3 MOBILE ROBOTIC MAPPING

Within the ﬁeld of mobile robotics a key concern is

providing the robot with the ability to acquire a model

of its operating environment as this model is required

for the safe and productiveoperation of the robot. The

actual performance of the robot in acquiring a mean-

ingful spatial model of its operating environment de-

pends greatly on its capability to quickly evaluate

the potentially erroneous information received from

its sensors. As it operates in the environment, the

robot gathers sensory information and subsequently

incorporates this into a representation of the envi-

ronment. Occupancy Grids have become the dom-

inant paradigm for environmental modelling in mo-

bile robotics because of their operational characteris-

tics (Kortenkamp et al., 1998) . The creation of these

Occupancy Grid maps is a non trivial process as the

robot has to interpret the ﬁndings of its sensors in or-

der to make deductions regarding the state of its en-

ICINCO 2010 - 7th International Conference on Informatics in Control, Automation and Robotics

vironment. This is facilitated by the use of a sensor

model which is a means of interpreting received sig-

nals through perceptual channels. In occupancy grid

based robotic mapping there are largely two types of

sensory model; the Inverse Model and the Forward

Model(Collins et al., 2007). In the context of our re-

quirements the inverse approach is currently most ap-

plicable. This is because it facilitates iterative real

time operation without any requirement for historical

knowledge and facilitates operation with low entropy

sensory information.

4 PHYSICAL ARCHITECTURE

A key design goal is to produce a portable system

which is able to run on constrained hardware. Al-

though the target device for the prototype is a mobile

phone it is envisaged that the software could run on

other more limited embedded devices. The overall de-

sign of the prototypecan be split into speciﬁc problem

domains involving obtaining the data, communication

of the data, encoding of the data and processing of the

data on the device itself.

4.1 Obtaining the Data

While movement will be detected by the accelerom-

eters contained within the device, an initial deci-

sion is how often to sample this movement and how

many samples are needed before we process the data.

Once we have accumulated sufﬁcient samples this

data needs to be analysed to decide whether or not we

think an adequate amount of shaking or movement is

taking place. Studies of accelerometer data include

calculating and using the covariance of the values ob-

tained (Ravi et al., 2005). For the prototype currently

in development we take the covariance of our X,Y

sample data using equation 1.

covar = (1/(n−1))

∑

i=1

− ˆx)(y

− ˆy) (1)

We then compare this covar result with a predeter-

mined threshold value. If it exceeds this threshold we

must then communicate our ﬁndings to other clients.

4.2 Communicating the Data

The ﬁrst challenge to overcome is deciding how to

broadcast or share data between multiple connected

devices in a scalable way. The Spread Toolkit offers

an open source solution based on a shared message

bus. It has been optimised to provide efﬁcient mes-

sage exchange with the ability to guarantee delivery

Determined Client

Confidence Store

Client Event Information

Individual Event Aspect

Dynamic Client Reference Table

Unique Client Reference

(Full Queue Depicted)

Dynamic Client confidence Table

Figure 1: Device architecture.

and ordering of messages if required. To improve

performance we will use unreliable communication.

Each mobile client connects to an individual Spread

daemon using a uniquely generated id. In addition,

Spread daemons can be connected together to form a

larger shared single communication bus where Spread

daemon 1 connects to Spread daemon n.

4.3 Processing the Data on the Device

All messages received by a single client will be

queued for a speciﬁc period of time. Thus, the number

of independent queues created reﬂects the number of

unique clients who have transmitted messages within

the sampling time period. Processing the queues in-

volves examining the number of queues at time t as

well examining their queue length. If the number of

queues is below a certain threshold or the mean queue

length is below a certain threshold we can reset all

queues and wait for the next sampling period before

repeating the process. Otherwise, we need to pro-

cess the data in the queues. For the prototype each

queue contains data representing a covariance value

v obtained from the accelerometer data. Each queue

will have an independent scalar value representing a

conﬁdence level k. Applying k(v

,.. .,v

) yields

(kv

,kv

,.. ., kv

) for each queue. Summation of

these queue vectors will provide a simplistic overview

of whether or not we suspect an earthquake is taking

place. Each time the queues are processed they are

cleared ready for the next sample. A key challenge

will be deriving an accurate conﬁdence scalar value

for each queue. This will ultimately need to take into

account historical data between sampling cycles.

All software used or written needs to be portable

and able to run on ARM and MIPS based hardware.

The software must also operate within the constraints

AN INVERSE SENSOR MODEL FOR EARTHQUAKE DETECTION USING MOBILE DEVICES

of the target device. The hardware used for the pro-

totype is the Openmoko Freerunner

. Signiﬁcant fea-

tures of the device include its accelerometers, WiFi

and GPRS. The ability to test over a GPRS connec-

tion will be important as it may not be possible to ac-

cess a wireless access point or there may not be a 3G

network available. Therefore being able to concisely

encode data for communication across Spread will be

essential. Regardless of the connection we also want

network communication to be light-weight in terms of

processing load. This eliminates standard approaches

such as structuring packets with XML data (Moore,

2007).

5 SEISMIC DETECTION MODEL

From an operational perspective the architecture we

are in the process of realising operates as follows.

Each device begins an operational cycle by populat-

ing its client event queue through taking in data prop-

agated from the various other devices in the network.

This per queue information must then be used as the

basis for determining the whether or not an event may

be happening. This problem is far from trivial as each

queue is subject to a potentially different and non de-

terministic sampling rate meaning that it is the in-

direct information contained within the queues that

must be used. In addition each device will have an in-

dependent view of the problem meaning that it is not

possible to directly rely on device interdependency

characteristics.

5.1 Client Queue Information

As a device can only determine information about the

operating environment indirectly through its sensor(s)

and the information propagated from its peer units the

determination of a world model is an applied exam-

ple of an estimation theory problem (Thrun, 2002).

Therefore to facilitate the interpretation of the data

provided from a client event a probabilistic sensor

model of the form p(r|z) is used. This model facil-

itates the derivation of the individual client event con-

ﬁdence values v, mentioned previously in section 4.

Therefore:

= p(r

)

where the model we use in this prototype is based

upon the characteristics outlined previously in section

4.3. This model relates the client event reading r to

the true event state z. This density function is sub-

sequently used in a Bayesian estimation procedure to

http://openmoko.com

determine the event state probabilities. Finally a de-

terministic world model is employed to facilitate the

derivation of a optimal world estimator which can be

propagated between the individual units that form the

world state.

A classical Bayesian approach is used for the deter-

mination of the per queue conﬁdence score. Given

the current estimate of the state of clientC

, p[s(C

) =

SE|{r}

] based on the observations r

= r

,.. ., r

and

given a new client observation r

t+1

the new state esti-

mate is provided by

k = p[s(C

) = SE|{r}

+ 1] = (2)

p[r

t+1

|S(C

) = SE]p[S(C

) = SE|{r}

]

∑

s(C

)

p[r

t+1

|s(C

)]p[s(C

)|{r}

]

In the above the previous estimated value of the client

state p[S(C

) = SE|{r}

] serves as the prior and is ob-

tained directly from a localised representation of the

global state. The new state of a particular client. de-

termined through the above, is subsequently stored in

this representation and propagated to the world.

To facilitate prior estimation for client state a simpli-

ﬁed one dimensional Gaussian estimator model is em-

ployed.

p(r|z) =

√

2πσ

exp



−(r−z)

2σ



(3)

5.2 Inter Device Conﬁdence Regions

As presented above the sensor model is a one di-

mensional construct associated with a determined or

evaluated distance between client devices. Therefore

the model can be considered a client information axis

from the one dimensional viewpoint. While useful

for determining information relating to the 1-1 spatial

mapping directly between the devices the model can-

not consider areas outside of this conceptual spatial

line. It is conceivable that the spatial area between a

host device and its client will is also an area of inter-

est. In particular it would be beneﬁcial to have the

ability to model a region of conﬁdence emanating di-

rectly from the host device to the immediate vicinity

of the client device.The basic premise of this concept

is outlined in ﬁgure 2. When extended in this manner

the probabilistic model approaches more closely the

type of robotic mapping inverse sensor model high-

lighted previously. The extended model can be spec-

iﬁed as equation 4 where Q is the angle associated

with the created conﬁdence region.

p(r|z, Q) =

2πσ

exp

−

(r−z)

2σ

(4)

ICINCO 2010 - 7th International Conference on Informatics in Control, Automation and Robotics

Client

Device

Distance of interest

from client device

Client information axis

(single dimension model)

Device 1

Confidence region

Figure 2: Creating a conﬁdence region between devices.

The availability of these inter device conﬁdence re-

gions will provide for a more information rich proﬁle

of the event to be computed. In addition the over-

lapping of such regions will provide for the ability of

assessing and verifying the information coming from

individual clients thus adding a novel dimension to

the client conﬁdence estimation.

6 MODEL CONSIDERATIONS

As the architecture evolves the consideration of sen-

sory units or other sources of relevant information is

necessary. These considerations are highlighted here.

6.1 Client Event Information

At its core the actual sensor model is a statistical es-

timation formulation which interprets relative range

information received from peer devices. Upon the ac-

tivation of a devices sensors an event signal is prop-

agated through the network. The determination of

realistic events on a device versus false positives or

false negatives is a separate problem to the data sig-

nal handling the sensor model is designed to consume

and hence an exposition of same is outside the scope

of this context. When an event signal is received the

sensor model calculates a probabilistic proﬁle for the

event. To illustrate, consider the ideal scenario where

a device receives notiﬁcation of an event from a peer

device at what is determined to be at distances of

60km and 100 km respectively from the device. The

associated probabilistic proﬁles determined through

the model are outlined in ﬁgure 3 where it can be

seen that the model is Gaussian in nature. In terms

of the device architecture each proﬁle corresponds to

a single component of an event queue. The preceding

example presented the model in the ideal scenario of

there being a 1-1 correspondence between the physi-

cal devices distance and the actual distance the infor-

mation has been determined to travel. In real world

settings such an assumption cannot be guaranteed.

The model takes cognisance of this fact by its nature

as illustrated in ﬁgure 4.

0.0000

0.1000

0.2000

0.3000

0.4000

0.5000

0.6000

0.7000

0.8000

0.9000

1.0000

0 20 40 60 80 100

Event Likelihood

Distance

Figure 3: Event proﬁles for hypothetical distances of 60km

and 100km respectively.

0.0000

0.2000

0.4000

0.6000

0.8000

1.0000

0 20 40 60 80 100

Event Likelihood

Distance

Peer device determined at 65km, actual 100km

Peer device determined at 50km, actual 60km

Figure 4: Event proﬁles in the context of non ideal peer

device distances. Hypothetical actual distances are 50km

and 65km respectively.

6.2 Information Source Integration

To increase the capability of any such system in gen-

eral requires that multiple sources of information can

be incorporated into a single, useful, information

source. This is known as the data fusion problem. Fu-

sion processes are frequently categorised as low, in-

termediate or high, depending on the processing stage

at which fusion takes place(Klien, ). Lowlevel fusion,

(Data fusion) combines several sources of raw data to

produce new raw data. The expectation is that fused

data is more informative and synthetic than the orig-

inal inputs. Within the context of multi-device seis-

mic detection this integration can be performed using

a formulation such as that outlined in equation 3 to

combine the estimates provided by the independent

clients. For two clients C

and C

this means using

the associated client data models p

(r|z) and p

(r|z)

as the basis for determining the associated combined

probability and subsequently applying an appropri-

ate normalisation across the state encapsulated in the

client conﬁdence table illustrated in ﬁgure 1.

7 CONCLUSIONS AND FUTURE

WORK

In this paper we have detailed a sensor modelling

framework for earthquake detection using mobile de-

AN INVERSE SENSOR MODEL FOR EARTHQUAKE DETECTION USING MOBILE DEVICES

vices which is used within the context of a novel seis-

mic event detection architecture. The sensor model

outlined is a probabilistic one Gaussian in nature and

similar to the inverse sensor models prevalent in the

robotic mapping ﬁeld. As such it is capable of in-

crementally and efﬁciently interpreting event signals

propagated throughout the network without the need

for predetermined models or sensor associated seg-

mentation decisions. For example the characterisa-

tion highlighted in section 6 illustrated that meaning-

ful client event evaluation is possible with a minimal

of information i.e. an event notiﬁcation and a client

distance estimate.

In terms of future work regarding the model and

its usage a number of areas are prevalent. The choice

of an inverse sensor model has some speciﬁc impli-

cations. Because of its theoretical basis the disam-

biguation and analysis of client event data is achieved

primarily through the use of additional sensing. This

has performance implications which need to be ad-

dressed. Another area of future work is determining

appropriatecharacteristics for the extension of the one

dimensional sensor model to two dimensions. The at-

tribute of interest here is determining a meaningful

distance of interest from a client device. To address

this problem we initially propose to employ simple

heuristic values determined from operational experi-

ence. Our long term aim however, is to facilitate the

automated derivation of the distance of interest, using

triangulation between clients. The evaluation of re-

ceived client events to determine the true likelihood

of an actual earthquake event as opposed to user di-

rected movement is another area of future research.

Benchmarking the detection ability of our technique

and subsequent model reﬁnement is also an obvious

area of future work. Toward this end we intend to cor-

relate our detection results with actual real earthquake

data obtained from national earthquake centres and

the Stanford Quake-Catcher Network. Finally within

the context of the project as a whole another impor-

tant area of future work will be the speciﬁcation of

a meaningful benchmarking technique, applicable to

the domain, to facilitate direct quantitative compari-

son between techniques such as ours and natural lan-

guage centric techniques such as the U.S. Geological

Surveys Twitter Earthquake Detector (TED)

REFERENCES

Collins, T., Collins, J., and Ryan, C. (2007). Occupancy

grid mapping: An empirical evaluation. In Proceed-

http://recovery.doi.gov/press/us-geological-survey-

twitter-earthquake-detector-t ed/

ings of Mediterranean Conference on Control and Au-

tomation.

Ehlers, F., Gustafsson, F., and Spaan, M. Signal processing

advances in robots and autonomy. EURASIP J. Adv.

Signal Process, 2009.

Hewitt, C. (1992). Open information systems semantics for

distributed artiﬁcial intelligence. Foundations of ar-

tiﬁcial intelligence Special Issue of ’Artiﬁcial Intelli-

gence’ Series, pages 79–106.

Klien, L. Sensor and data fusion: A tool for information

assessment and decision making. SPIE Press.

Kortenkamp, D., Bonasso, R., and Murphy, R. (1998). AI-

based Mobile Robots: Case studies of successful robot

systems.

Moore, J. P. T. (2007). Thumbtribes: Low bandwidth,

location-aware communication. In Obaidat, M. S.,

Lecha, V. P., and Caldeirinha, R. F. S., editors, WIN-

SYS, pages 197–202. INSTICC Press.

Murooka, T., Takahara, A., and Miyazaki, T. (2001). A

novel network node architecture for high performance

and function ﬂexibility. In ASP-DAC, pages 551–557.

Newmark, N. and Rosenblueth, E. Fundamentals of earth-

quake engineering. Prentice-Hall.

Ravi, N., Dandekar, N., Mysore, P., and Littman, M. L.

(2005). Activity recognition from accelerometer data.

In IAAI’05: Proceedings of the 17th conference on In-

novative applications of artiﬁcial intelligence, pages

1541–1546. AAAI Press.

Thrun, S. (2002). Robotic mapping: A survey. In Lake-

meyer, G. and Nebel, B., editors, Exploring Artiﬁ-

cial Intelligence in the New Millenium. Morgan Kauf-

mann.

Vargas, F., Fagundes, R. D., and D. Barros, J. (2001). Sum-

marizing a new approach to design speech recognition

systems: A reliable noise-immune hw-sw version. In-

tegrated Circuit Design and System Design, Sympo-

sium on, 0:0109.

Zareian, F. and Krawinkler, H. (2009). Simpliﬁed perfor-

mance based earthquake engineering. Technical re-

port, Stanford University.

ICINCO 2010 - 7th International Conference on Informatics in Control, Automation and Robotics