JOURNAL OF E 11, VOL. 1, NO. 1, AUGUST 2009 1

Community Seismic NetworkDaniel Obenshain, K. Mani Chandy, Rishi Chandy, Rob Clayton, Andreas Krause, Michael Olson,

Daniel Rosenberg, and Annie Tang

Abstract—Geologists have long sought an early warning systemfor earthquakes since it is the best way to save lives and preventproperty damage. Currently, the seismic network in the LosAngeles area is too sparse to provide any early warning. Wepropose a distributed system using MEMS sensors attached tovolunteers’ computers in homes and schools across the LosAngeles area. This system will have the advantage of a fasterresponse time and a denser network than the existing system.

Index Terms—distributed computing, seismic network.

I. INTRODUCTION

EARTHQUAKES are devastating natural disasters. Theyare of particular concern to the people of the Los Angeles

area as the San Andreas Fault and other faults lie nearby.ShakeOut, a scenario constructed by the United States Geo-logical Survey, conservatively estimates that a 7.8 or highermagnitude earthquake in this area would result in a death tollof 2,000, with an additional 50,000 injured and over $200billion in damage[4].

In such an event, early warning would be a lifesaver. With aslittle as ten seconds of early warning, automated systems couldtake action to reduce damage and loss of life. Elevators couldbe stopped and the doors opened so that the passengers couldstep out. Data servers could suspend read/write operationsto prevent corrupting millions of dollars worth of data. Theelectrical grid could be placed in a more stable configurationto reduce rolling blackouts as a result of downed wires (fromconversations between Dr. K. Mani Chandy and SouthernCalifornia Edison).

Early warning systems are already in place in Japan, Mex-ico, Romania, Taiwan, and Turkey[1]. At the University ofCalifornia, Riverside scientists are also working to distributeseismic sensors to volunteers[2]. In this paper, we describea community-based sense and response system using a dis-tributed network of sensors. The network proposed is differentfrom most of the earlier work in its emphasis on getting largenumbers of volunteers from the community to buy and installinexpensive sensors or to use sensors in their mobile phonesand laptops, and also to participate in responding to warnings.

In our system, a volunteer will be able to use one of severaldifferent kinds of accelerometers to make his or her computera client in our network. Each client will then log seismicdata using the accelerometer. If significant shaking occurs, theclient computer will “pick” that data and then send a messageto the server, alerting the server to possible earthquake activity.

The server will receive a stream of picks from clientsscattered all over the Los Angeles area. The server will then

Daniel Obenshain would like to thank Caltech and the SURF Program,especially Dr. Kiyo and Eiko Tomiyasu, who sponsored his SURF.

Manuscript received September 25, 2009.

evaluate the incoming picks to determine if it is likely thanan earthquake is occurring and, if so, where it is. The serverwill then immediately generate a ShakeMap[6], which willbe very useful in evaluating damage and organizing reliefefforts. Optimally, an estimate of the size and location ofthe earthquake will be generated before the shaking actuallyoccurs, allowing the system to distribute that information inthe form of an early warning.

The data-picks stream will contain a lot of noise. Since thesesensors will not be underground or connected to bedrock, theywill be subject to the vibrations in their environments. It isreasonable to assume that users will often accidentally set offtheir sensors by bumping them or kicking the table. Becauseof this, it is important to have a dense network so that errorswill be dampened by the information from the surroundingnetwork.

Since validation of the system is very important, the networkwill have an additional playback feature. If requested, theserver can distribute acceleration data to the clients to havethem play back. Then, at a specified start time, the clientswould play back the data to simulate an event across the entirenetwork. This should greatly increase public confidence in thesystem.

II. SYSTEM DESIGN

A. Conceptual

This network of sensors will rely on Internet communica-tions and volunteer support from the public. Each participat-ing individual will either purchase an accelerometer device(MEMS or similar) and connect that device to his or hercomputer or use an existing accelerometer device inside thecomputer. The participant will then download our softwarefrom our server and execute the software to begin sendingdata. Once the data is received by the server, the server willbe able to use that information for early warning.

B. Client

Once the client has downloaded and installed the software,the client will be part of our seismic network. Whenever theclient’s computer is free, it will read data periodically fromthe accelerometer and store it in a ring buffer (a buffer whereadding a new value deletes the oldest existing value if thebuffer is full) on the client’s computer. The ring buffer willbe read by a picking algorithm, which will trigger if seismicactivity is suspected and send a message to the server. Inorder to minimize the inconvenience to the client, the wholeapplication should take less than one percent of the system’sresources.

JOURNAL OF E 11, VOL. 1, NO. 1, AUGUST 2009 2

Fig. 1. The client is able to see the data gathered by his or her computer.

To encourage use, the client will be able to watch the datagathered by his or her machine in real-time via a graphicaluser interface (Figure 1).

In addition to these features, the client will have the abilityto “play back” old data. Every day, the client will “call home”via a heartbeat message to alert the server to its presence.During this exchange, the client will send a message to theserver notifying the server that the client is functioning. Forsecurity purposes, the server will never contact any client;only the client may initiate contact. After receiving the initialmessage, the server will update the client’s parameters, alertthe user to an improved version of the software if one exists,ask for any logs saved by the client, and alert the client to anyplayback requests. If the client receives a playback request, itwill create a new thread which will then wait until the specifiedstart time. The new thread will then start reading data fromthe file given by the server, rather than the accelerometer. Thiswill allow us to test the network from end to end.

C. Server Architecture

The server will collect all the pick data sent by the clients. Itwill then use existing Associator and Locator code to pinpointthe earthquake event. It will output a ShakeMap from this datato aid in early warning and relief efforts.

III. ALGORITHMS

A. Signal Processing

The acceleration data is gathered by either a MEMS deviceattached to the computer, or an internal accelerometer device.With proprietary drivers, that data is gathered by the clientand stored in a ring buffer (a buffer where the oldest value isdeleted to make room for new values if the buffer is full).

To normalize the values, an average is calculated over eachaxis of the accelerometer and data is reported relative to thisaverage. Gravity is discounted, since the z-axis normal valuewill account for this, so the client will only deal with changesfrom this normal value.

B. Picker

Optimally, the algorithm will “pick” acceleration data ifand only if seismic activity is present. More practically, thealgorithm will “pick” whenever there is a detectable changein the acceleration detected by the client.

The picking algorithm depends on a short term averageof recent data points and a long term average of recentdata points. The number of points aggregated by averageis variable, but currently the short term average aggregatesthe most recent ten data points, while the long term averageaggregates the most recent 250 data points. Each representsthe average absolute magnitude acceleration seen by the clientover all three spatial dimensions, minus the normalizationvalues to discount the steady state. If the short term averageexceeds the long term average by a threshold value, the clientassumes that seismic data is occurring and “picks” that data.The threshold value is variable, to account for different sensorconditions for different clients, but a ten percent thresholdvalue would be reasonable.

Once the picking algorithm has triggered, the software willimmediately save the contents of the ring buffer to the harddrive on the assumption that the computer might soon losepower. Then, the software waits some time tpause to determinethe largest magnitude acceleration seen in that time. Thesoftware then packages that information in SAC format andsends it to the server via UDP, to avoid waiting for a responsefrom the server. Then, some time later, the software saves thering buffer to the hard drive again to ensure that the wholeevent was captured.

The client then waits some time tdelay before pickingagain, to avoid picking multiple times during the coda of theearthquake.

Both tpause and tdelay are parameters that can be tuned,both on the network level and at the level of the individualclient. If tpause is too long, the shaking information will notreach the server in time and the server will not be able to givean early warning. On the other hand, if tpause is too short,the client will report an acceleration value that is too small,which will result in incorrect estimates of the earthquake’smagnitude. If tdelay is too short, the client will send severalmessages to the server for the same set of shaking, which willbog down the server. If tdelay is too long, the client mightmiss a second earthquake occurring soon after the first one.The server will be able to update these values on the client’scomputer via the heartbeat process to make the network moreeffective.

C. Associator

The server will receive a constant stream of data from theclients and the clients “pick” data. An Associator program onthe server will associate the incoming data with existing datato determine if it is likely that an earthquake has just occurred.

If such an earthquake is likely, the server will determinethe likely source and intensity of the earthquake and pass onthis information to interested parties. Ideally, this informationcould then be distributed before significant shaking had oc-curred.

JOURNAL OF E 11, VOL. 1, NO. 1, AUGUST 2009 3

Fig. 2. The accelerometer device, before it has been connected to a computer.

IV. IMPLEMENTATION

The client program was written in Java to provide a platformindependent program. An additional C library provided bySkyhook Wireless was included to locate the clients, usingSkyhook Wireless software.

In the operation of the system, there are three types ofmessages exchanged between the client and the server. Whenthe client is initialized, it will send a registration message tothe server with its current location and be given a unique clientID in return. When the client undergoes significant shaking, itwill send a pick message with a record of that shaking to theserver. Periodically, the client will also “call home” to requestany updates. The server will respond with a list of parameterchanges, any code updates, any requests for saved data, andany playback requests.

If the server responds with playback requests, the client willsave that data and wait until the specified start time. Then, itwill start a new thread where it will read in the saved data asif from an accelerometer. This will allow us to test the systemfrom end to end with either recorded or synthetic data.

V. EVALUATION

Since it is vitally important to establish the reliability ofour sensors, we tested one of the sensors against an existingsensor in the Southern California Seismic Network (SCSN).

We set up our sensor (Figure 3) in the basement of MillikanLibrary on Caltech’s campus, next to a conventional sensorMIKB (Figure 4). Both devices were placed in a smallbasement maintenance room (Figure 5). After simulating anearthquake by hitting the concrete floor with a sledgehammer,we compared the waveforms received by both devices.

While the data gathered by the MEMS sensor had muchmore noise than the conventional sensor, both detected accel-eration spikes at the same times. With more sensors, it will bepossible to overcome the additional noise with a dense networkof sensors.

VI. CONCLUSIONS

Here we have shown a plan to create a dense networkof cheap sensors to detect earthquakes. We have shown thatour sensors can detect seismic activity, though not as well asconventional sensors. We have outlined the pattern of network

Fig. 3. An example client. The accelerometer device is seen here connectedto the USB port on an example laptop.

Fig. 4. Sensor MIKB. This is one of the sensors in the Southern CaliforniaSeismic Network, and is designated MIKB since its location is in the basementof Millikan Library on the Caltech campus.

Fig. 5. The arrangement of the two sensors. Both the example client andthe MIKB sensor are visible. Also seen here is the sledgehammer used tosimulate seismic activity for testing purposes.

Fig. 6. The impact of the sledgehammer, as detected by the two sensors.The signal detected by our sensor is much less clean.

JOURNAL OF E 11, VOL. 1, NO. 1, AUGUST 2009 4

traffic between the server and client and shown how we planto implement a dense network of sensors.

VII. FURTHER WORK

In the future, we look forward to expanding this project toinclude mobile devices, such as cell phones and laptops withbuilt-in accelerometers.

We also look forward to implementing this in the field,which will give us some real-life data.

ACKNOWLEDGMENT

Daniel Obenshain would like to thank Dr. Carol Readheadand others for proofreading this manuscript.

REFERENCES

[1] R. Allen and H. Kanamori, “The potential for earthquake early warningin southern California.” vol. 300: American Association for the Advance-ment of Science, 2003, pp. 786-789.

[2] E. Cochran, J. Lawrence, C. Christensen, and R. Jakka, “The Quake-Catcher Network: Citizen Science Expanding Seismic Horizons,” Seis-mological Research Letters, vol. 80, p. 26, 2009.

[3] G. Cua and T. Heaton, “‘7 The Virtual Seismologist (VS) Method: aBayesian Approach to Earthquake Early Warning,” Earthquake EarlyWarning Systems, p. 97, 2007.

[4] L. Jones, R. Bernknopf, D. Cox, J. Goltz, K. Hudnut, D. Mileti, S. Perry,D. Ponti, K. Porter, and M. Reichle, “Earthquake Hazards Program,” USGeological Survey Open File Report, p. 1150, 2008.

[5] D. H. Oppenheimer, A. N. Bittenbinder, B. M. Bogaert, R. P. Buland, L.D. Dietz, R. A. Hansen, S. D. Malone, C. S. McCreery, T. J. Sokolowski,P. M. Whitmore, and C. S. Weaver, “The seismic project of the NationalTsunami Hazard Mitigation Program,” Natural Hazards, vol. 35, pp. 59-72, May 2005.

[6] D. Wald, V. Quitoriano, T. Heaton, H. Kanamori, C. Scrivner, and C.Worden, “ “TriNet ShakeMaps”: rapid generation of peak ground motionand intensity maps for earthquakes in southern California,” EarthquakeSpectra, vol. 15, pp. 537-556, 1999.

K. Mani Chandy is the Simon Ramo Professor atthe California Institute of Technology.

Dr. Chandy got his Ph.D. from the MassachusettsInstitute of Technology in Electrical Engineering atthe Operations Research Center in 1969. He got aMasters from the Polytechnic Institute of Brooklyn,and a Bachelors from the Indian Institute of Tech-nology, Madras in 1965.

Dr. Chandy has worked for Honeywell and IBM.From 1970 to 1989, he was in the Computer ScienceDepartment of the University of Texas at Austin,

serving as chair in 1978-79 and 1983-85. He has served as a consultantto a number of companies including IBM and Bell Labs. He has been atthe California Institute of Technology since 1987, two years as a ShermanFairchild Fellow, and then as a professor in Computer Science.

Dr. Chandy is a member of the National Academy of Engineering. Hereceived the IEEE Koji Kobayashi Award for Computers and Communicationin 1987, the A.A. Michelson Award from the Computer Measurement Groupin 1985, and has numerous awards.

Software developed by Dr. Chandy and colleagues in the area of computerperformance modeling was marketed by Boole and Babbage Inc.. He was aco-founder of a company, iSpheres, in the area of event-driven architecture;that software is now marketed by Avaya.

Dr. Chandy does research on sense and respond systems. He has publishedthree books and over a hundred papers on distributed computing, verificationof concurrent programs, parallel programming languages and performancemodels of computing and communication systems.

Rishi Chandy is a junior majoring in ComputerScience at the California Institue of Technology.

Rob Clayton B.A.Sc., University of Toronto, 1973;M.Sc., University of British Columbia, 1976; Ph.D.,Stanford University, 1981. Assistant Professor of Ex-ploration Geophysics, Caltech, 1981-85; AssociateProfessor, 1985-89; Professor of Geophysics, 1989-.Executive Officer for Geophysics, 1987-94; ActingDirector, Seismological Laboratory, 1989; 2008-09;Deputy Director, 1989-90; Academic Officer, 2008-.

Andreas Krause Assistant Professor of ComputerScience California Institute of Technology

Ph.D. Computer Science, Carnegie Mellon Uni-versity, 2008 Dipl.-inf., Dipl.-math., Technische Uni-versitt Mnchen, Germany, 2004

Daniel Obenshain is a senior at the California Insti-tute of Technology. He is currently double majoringin Computer Science and English.

Michael Olson is a graduate student at the Califor-nia Institute of Technology working under Profes-sor Mani Chandy researching applications of event-driven architectures. Current research explores fed-erated systems of Internet applications as a means toidentify the most relevant content in a given area.

Michael got his B.S. from Carnegie Mellon Uni-versity in Business Administration and ComputerScience in 2004. From 2004 to 2007, Michaelworked for Deloitte Consulting as an SAP consul-tant, providing advice and implementation assistance

for the finance module to five different clients.

JOURNAL OF E 11, VOL. 1, NO. 1, AUGUST 2009 5

Daniel Rosenberg is a junior at the CaliforniaInstitute of Technology.

Annie Tang is a sophomore at the California Insti-tute of Technology.