Hybrid Seismic Response Simulation on a GeographicallyDistributed Bridge Model
Gilberto Mosqueda1; Bozidar Stojadinovic2; Jason Hanley3; Mettupalayam Sivaselvan4; andAndrei M. Reinhorn5
Abstract: The hybrid simulation test method is a versatile technique for evaluating the seismic performance of structures by seamlesslyintegrating both physical and numerical simulations of substructures into a single test model. Using hybrid simulation, the seismicresponse of complex structural systems partitioned into multiple large-scale experimental and numerical substructures at networkedgeographically distributed experimental and computational facilities can be evaluated. A scalable framework with a fault-tolerant distrib-uted controller is presented to support the implementation of advanced hybrid testing methods with distributed substructures. The controlstrategy is based on a multithreaded simulation coordinator for parallel communication with remote sites and an event-driven controllerat each remote experimental site to implement continuous loading. The distributed controller provides faster rates of testing and improvedaccuracy in the simulation results. The effectiveness of the proposed framework is demonstrated by computing the earthquake response ofa six-span bridge model with multiple remote experimental and numerical column substructures distributed within NEES laboratoriesacross the United States. Further, the distributed tests were implemented using a secure network link between the testing sites that wasdeveloped for the NEES cyber infrastructure.
Hybrid simulation combines numerical and experimental methodsto evaluate the seismic performance of structures. The principlesof the hybrid simulation test method are rooted in the pseudody-namic testing method developed over the past 30 years �Taka-nashi et al. 1975; Takanashi and Nakashima 1987; Mahin et al.1989�. In a hybrid simulation, the dynamic equation of motion issolved for the hybrid numerical and experimental model. Typi-cally, the experimental substructures are portions of the structurethat are difficult to model numerically, thus, their response ismeasured in a laboratory. Numerical substructures represent
1Assistant Professor, Dept. of Civil, Structural and EnvironmentalEngineering, State Univ. of New York, Buffalo, NY 14260-4300�corresponding author�. E-mail: [email protected]
2Associate Professor, Dept. of Civil and Environmental Engineering,Univ. of California, Berkeley, CA 94720-1710.
3Information Technology Service Manager, Structural Engineeringand Earthquake Simulation Laboratory, State Univ. of New York, Buffalo,NY 14260-4300.
4Assistant Professor, Dept. of Civil, Architectural and EnvironmentalEngineering, Univ. of Colorado, Boulder, CO 80309.
5Clifford C. Furnas Professor of Structural Engineering, Dept. of Civiland Environmental Engineering, State Univ. of New York, Buffalo, NY14260-4300.
structural components with predictable behavior: they are mod-eled using a computer. Comparative experiments have shown thatthe hybrid simulation test method in which inertial effects aresimulated numerically can provide results comparable to theshaking table test method when propagation of experimental er-rors is successfully mitigated �Takanashi and Nakashima 1987;Mahin et al. 1989; Magonette and Negro 1998�.
Hybrid simulation procedures based on the pseudodynamic ap-proach have advanced considerably since the method was firstdeveloped. Early tests utilized a ramp-and-hold loading procedureon the experimental elements. Continuous testing at slow �Ma-gonette 2001� and fast rates �Nakashima et al. 1992; Nakashima2001� have improved hybrid simulations by mitigating strain-raterelated errors �force relaxation� during displacement hold portionsof the tests �Stojadinovic et al. 2006�. The capabilities of hybridsimulation have been further extended by proposing to geographi-cally distribute experimental substructures within a network oflaboratories linked through numerical simulations using the inter-net �Campbell and Stojadinovic 1998�. In this approach, themaster simulation solving the equation of motion of the entirestructure can be carried out on supercomputers if necessary andnetworked to one or more experimental facilities with remotesubstructures. The infrastructure of the George E. Brown Jr. Net-work for Earthquake Engineering Simulation �NEES� providesthe experimental facilities, the numerical modeling tools, and thenetwork interface to enable simultaneous testing of multiplelarge-scale experimental and numerical substructures using dis-tributed hybrid simulation. This approach allows for the evalua-tion of complex structural systems through the partitioning of theof the model into multiple experimental and numerical substruc-tures. The experimental substructures can be evaluated at large or
full scale under realistic simulated seismic loading while at the
same time capturing the interaction of the test specimens with thecomplete structural system.
Geographically distributed hybrid simulations have been car-ried out jointly between Japan and Korea �Watanabe et al. 2001�,in Taiwan �Tsai et al. 2003� and in the United States �Spenceret al. 2004a; Mosqueda et al. 2005; Stojadinovic et al. 2006�.More recently, researchers have been investigating the use of ad-vanced numerical simulation tools �Takahashi and Fenves 2006;Wang et al. 2005� and implicit integration algorithms �Pan et al.2005� to improve the numerical simulation capabilities for hybridsimulation with geographically distributed substructures. The dis-tributed testing methods presented here can be integrated intothese various simulation platforms to minimize experimental er-rors and achieve reliable hybrid simulations of complex structuralsystems.
The objectives of the distributed control strategy presentedhere are: �1� to provide a scalable framework for multiple sub-structure testing at distributed sites; �2� improve the reliability ofthe test results by minimizing strain rate and force relaxationerrors in the remote experimental substructures; and �3� increasethe speed of testing as allowed by network communication time.The control strategy is based on a multithreaded simulation coor-dinator combined with an event-driven controller at the remoteexperimental sites. The multithreaded coordinator is applied tosimultaneously load multiple remote substructures at differentsites. The event-driven remote site controller allows for the imple-mentation of continuous hybrid simulation algorithms on distrib-uted models where computation, network communication, andother tasks may have random completion times. The advantage ofthis approach is that the hold phase in conventional ramp-and-hold pseudodynamic testing, during which force relaxation hasbeen observed, is minimized if not eliminated. These combinedfeatures also allow for faster rates of testing.
The distributed controller presented here was implementedinto NEESGrid, the cyber infrastructure linking the NEES equip-ment sites through and experiment-based deployment activity ofthe NEES system integration involving the University of Califor-nia at Berkeley, University at Buffalo, University of Colorado atBoulder, the University of Illinois at Urbana-Champaign, and Le-high University. This combined effort known as fast multisiteon-line simulation test �Fast-MOST�, was targeted at introducingfeatures into NEESGrid �Spencer et al. 2004b� that allow forfaster rates of testing and improved reliability of the simulationresults. Building on the original MOST �Spencer et al. 2004a�,distributed control strategies were implemented into the NEEStele-control protocol �NTCP� �Pearlman et al. 2003� in order toincrease the speed of testing and allow for the implementation ofcontinuous algorithms.
The effectiveness of the distributed control system is demon-strated by computing the earthquake response of a six-span bridgemodel with five experimental and numerical column substructuresdistributed within NEES facilities. Although five NEES equip-ment sites were involved in this effort, the distributed simulationspresented here involve only two experimental substructures atBerkeley and Buffalo. The proposed distributed control strategywas implemented using NEESGrid to provide a secure networklink between the NEES equipment sites. Results from these simu-lations are presented, including a summary of task times andproblems encountered in network communication, computational
operations, and experimental loading of physical substructures.
The purpose of Fast-MOST was to combine the state of the art inhybrid testing with the state of the art in secure network commu-nications. Use of the NTCP network protocol in hybrid simulationwas first demonstrated in the July 2003 MOST �Spencer et al.2004a�. In this previous application, the emphasis of NTCP wason security and safe control of remote experimental equipment. Inorder to increase the rates of testing for Fast-MOST, three keyenhancements were incorporated into NTCP: �1� modification ofNTCP to minimize network transactions in each simulation step;�2� implementation of a Java-based multithreaded simulation co-ordinator to carry out transactions in parallel with multiple remotesites; and �3� implementation of an event-driven controller at re-mote experimental sites that generate a continuous load historyfor the experimental substructures. It should be noted that inmodifying the original NTCP, control and security features had tobe relaxed. Thus a balance was sought between the most essentialsecurity and control features that would provide the fastest ratesof testing. Detailed implementation of performance enhancementsis discussed in the sections that follow.
NTCP Improvements
The original NTCP protocol used in MOST was designed forsecurity and reliability �Pearlman et al. 2003�. Design goals wereto provide a mechanism for conducting multisite distributed test-ing using a standard, well defined protocol. The resulting protocolprovided effective safety control features to pause and restart thesimulation in case problems were encountered, but was less effi-cient in terms of network utilization and the rate of testing. Typi-cal integration step durations during the MOST experiment wereon the order of 13 s with the majority of this time dedicated tonetwork communication and overhead in the software interface.
For each integration step in the MOST simulation, at least tworound-trip network communications were required for each re-mote site. First, the simulation coordinator received the targetdisplacements from the master simulation. Next, the simulationcoordinator executed a propose request with the target encapsu-lated in a control point parameter for each remote site. Each re-mote site replied back to the simulation coordinator indicatingwhether it accepted its proposal. If the proposed request was ac-cepted for all sites, the simulation coordinator sent an executerequest to each site. Upon receiving an execute request, the ex-perimental sites commanded the actuator to the target displace-ment and returned the measured displacement and forces to thesimulation coordinator. Finally, the simulation coordinator sentthe feedback to the master simulation and repeated the process inthe next step.
For the Fast-MOST experiment, a new NTCP command Pro-poseAndExecute was created, which combined propose andexecute into a single command. For each step, the ProposeAndEx-ecute command is called with the control point parametercontaining the target displacement. The remote site responds bycommanding the actuator to the target then returning the force anddisplacement feedbacks to the master simulation. This addition tothe protocol reduces the number of round-trip communications toone per site in each simulation step. As a result, network commu-nication time can be reduced by half for the distributed test. TheProposeAndExecute command required changes to both theNTCP server and client interfaces but remained backward com-patible with existing client and control plug-ins �Pearlman et al.
2004�. Control systems interfaced with NTCP can take advantage
134:535-543.
Dow
nloa
ded
from
asc
elib
rary
.org
by
Uni
vers
ity o
f B
righ
ton
on 0
6/27
/14.
Cop
yrig
ht A
SCE
. For
per
sona
l use
onl
y; a
ll ri
ghts
res
erve
d.
of this new command since it was implemented on the server andrequires no changes to the control plug-in interface.
Multithreaded Communication with Multiple Sites
The simulation coordinator plays a central role in a distributedexperiment and acts as an interface to transfer data between themaster simulation and the remote sites. As an NTCP client,the simulation coordinator communicates with the remote sitesrunning NTCP servers over the network. The NTCP servers atremote sites are interfaced with local actuator controller and dataacquisition systems. The reference simulation coordinator used inthe MOST experiment was implemented in Matlab and commu-nicates with the master simulation via NTCP. The Matlab versionof the simulation coordinator imposes a number of performancelimitations that add significant overhead to the execution of a test.The NTCP client interface is written in Java and accesses Matlabthrough a wrapper interface that induces delays and overhead intoeach call to NTCP. Also, Matlab allows for only single threadedapplications, requiring serial communication to each site.
To overcome these limitations, a multithread simulation coor-dinator, shown in Fig. 1, was developed in Java and is called theJava Simulation Coordinator. For improved efficiency, the JavaSimulation Coordinator and master simulation were combinedinto the same program as different software modules. Thisapproach eliminated the network communication between thesimulation coordinator and the master simulation required in theoriginal MOST experiment. Further, the elimination of the sepa-rate master simulation reduces the number of required round-tripnetwork communications per step by one. To maintain versatility,the master simulation module was designed to be easily replacedby other simulation packages through a simple interface. Princi-pally, communications with the remote sites is parallelized usingseparate concurrent threads in the new simulation coordinator.This allows the time for each step to be equal to the time taken bythe slowest remote site, instead of the sum of the time taken by allsites. This implementation also uses the ProposeAndExecute com-mand to further reduce the number of network communications
Fig. 1. Distributed control architecture for hybrid simulation used inMOST and Fast-MOST
with each site.
JOUR
J. Struct. Eng. 2008.
Event-Driven Distributed Controller
Continuous testing, as opposed to conventional pseudodynamictesting using ramp-and-hold procedures to load experimental sub-structures, can minimize force relaxation and strain-rate relatederrors in hybrid simulations �Stojadinovic et al. 2006�. The chal-lenge in applying continuous methods for multisite tests is thatthey are based on real-time algorithms. In a distributed testingenvironment involving the internet, network communication timeis random and the integration task time may be random if notimplemented on a real-time platform. A fault-tolerant system isneeded to deal with random tasks, particularly for internet com-munication where network delays are likely to occur. Addition-ally, the distributed simulation should be protected against rareevents, such as the integrator crashing or loss of communicationwith remote sites. In such events, the simulation results can besalvaged and the tests continued after recovery.
In cases where task execution times are random, a clock-basedcontrol scheme could fail if the required processes are not com-pleted within the allotted time. As an alternative to the clock-based scheme used for real-time applications, an event-drivenreactive system, based on the concept of finite state machines�Harel 1987�, can respond to events based on the state of thehybrid simulation. The event-driven system can be programmedto account for the complexity and randomness of real systems inways that minimize the random effects on experimental substruc-tures. The programming procedure is based on defining a numberof states in which the program can exist and the transitions be-tween these states as events occur.
The state transition diagram in Fig. 2 shows the implementa-tion of an event-driven version of Nakashima and Masaoka’s�1999� polynomial approximation method. This algorithm con-tinuously updates the actuator commands using polynomial ex-trapolations of known displacement values to predict the actuatorcommands and interpolates once the target displacement is knownto approach the correct target displacement. The state diagramconsists of five states: extrapolate, interpolate, slow, hold, andfree�vibration. The default state is extrapolate, during which thecontroller commands are predicted based on previously computeddisplacements. The state changes from extrapolate to interpolateafter the controller receives the next target displacement and gen-
Fig. 2. Event-driven controller running modified version ofNakashima and Masoaka’s �1999� polynomial interpolation/extrapolation algorithm
erates the event D�update. The event D�target is generated once
the physical substructure has realized this target displacement.The model then subsequently transitions back to the extrapolatestate and sends updated measurements to the integrator. Thesmooth execution of this procedure is dependent on selectingthe run time of each integration step sufficiently large for all ofthe required tasks to finish. Small variations in completion timesfor these tasks will only affect the total number of extrapolationsteps versus interpolation steps.
Network Architecture
Fig. 3 shows the typical geographically distributed simulationsetup including the master site, running the simulation coordina-tor, and one remote site where a substructure �physical or numeri-cal� is located. The connection between the two sites is throughInternet2, an advanced research and education network �Internet22005�. The master simulation connects to the NTCP server run-ning at the remote NEES site. The NTCP server is customized foreach site with a control plug-in that serves as an interface to thelocal actuator control and data acquisition system. This controlplug-in communicates with the controller and relays all com-mands and feedback to the NTCP server.
The Java version of the simulation coordinator and mastersimulation, coupled with the improvements to NTCP, reduced thenumber of total round-trip communications to one per step foreach remote site. This is in contrast to the six network communi-cations executed per step for the MOST experiment, which had atotal run time of 5.5 h and an average step time of 13.2 s. Reduc-ing the network communications overhead, the average step timefor the Fast-MOST was expected to be 2 s. Further reductions inthe time required to move the actuators to the target displace-ments were also expected based on the predictors used in theevent-driven controller at remote sites.
Scalable Framework
A useful feature of the network architecture and NTCP used hereis scalability. This allows users to add more numerical and experi-mental substructures and readily replace experimental substruc-tures with numerical models �Pearlman et al. 2004�. With themultithreaded simulation coordinator described previously, addi-tional remote substructures can be included in a simulation with-out significantly increasing the simulation time. In order toparticipate in a simulation, a laboratory needs a “remote site”setup as shown in Fig. 3. The main prerequisite is a PC with the
Fig. 3. Network interface between ma
NEESpop software including NTPC and a control plug-in, which
forwards the NEESpop commands to the control system. The con-trol plug-in is custom to each site to adapt to different hardware.
The scalable framework also provides the ability to conductpretest simulations where the experimental substructures are re-placed by numerical substructure models. Further, the numericalsubstructures can be local to the master simulation or can becomputed at a remote site to include the effects of network com-munication in the simulation. The software switches to changebetween these options allows for the step-by-step setup and veri-fication procedure of the network interface and communicationprotocols for the experimental hardware. Remote sites can be in-troduced in the simulation one at a time to detect errors. Thefollowing pretest sequence is suggested to ensure that all sitesinvolved in a multisite test are on-line and working correctly:1. All sites conduct independent local simulations to ensure that
the software and interface to the experimental setup is func-tioning, particularly the NEESpop control plug-in used tocommunicate with the control system as show in Fig. 3.Communication between the master site running locally andthe NEESpop shown in Fig. 3 can utilize the local area net-work and Ethernet to simulate the Internet;
2. Conduct a distributed numerical simulation involving all par-ticipating sites as shown in Fig. 1. In this step, the remoteexperimental sites should use a numerical simulation of theexperimental setup. This step verifies the communicationlink between the NTCP servers and the NTCP client at themaster site. In addition, the simulations conducted as part ofthis step will provide data on the network timing required perstep and characterization of network delays;
3. Conduct a distributed simulation replacing one numericalsimulation with the actual experimental setup. Incrementallyadd in one actual experimental site at a time and verify per-formance. If using physical specimens, these tests should beconducted at low excitation amplitude to avoid damaging thespecimens. Alternatively, a numerical simulation of the ex-periment can be used while maintaining the control systemhardware in the loop; and
4. Finally, conduct the planned hybrid simulation after all thenetwork and interfaces to the experimental setups have beenverified.
Based on past experience in conducting distributed simula-tions, pretest numerical simulations without physical specimensare important to identifying and resolving problems associatedwith interfacing different hardware components. In particular, twoseparate models have been useful: one of the control system andexperiment combined, and a second model of the experiment onlybut physically exercises the actual control system hardware. The
mulation and remote substructure site
ster si
first model is especially useful in Step 2 while working out the
134:535-543.
Dow
nloa
ded
from
asc
elib
rary
.org
by
Uni
vers
ity o
f B
righ
ton
on 0
6/27
/14.
Cop
yrig
ht A
SCE
. For
per
sona
l use
onl
y; a
ll ri
ghts
res
erve
d.
details of the internet communication, and the second model isnecessary to test the communication between the NEESpop andcontroller in Step 3. A similar approach to modeling control sys-tems and experimental setups has been useful in the developmentand real-time verification of active, hybrid, and semiactive struc-tural control algorithms �Chu et al. 2005�.
Multisite Test
A multisite hybrid simulation was carried out to demonstrate theproposed distributed controller with teleoperation capabilitiesusing NEESGrid. For this purpose, a simple bridge model wasselected and substructures consisting of the bridge columns weredistributed as experimental substructures to participating NEESsites. A description of the structural model, including the experi-mental substructures and the numerical algorithms used to evalu-ate the earthquake response, is provided next. In the sections thatfollow, the preparation steps and resulting data from an actualsimulation are presented.
Structural Model
A six-span bridge with five columns, as shown in Fig. 4, wasselected as the structural model for the Fast-MOST experiment.The dimensions and element properties of the structural systemare loosely based on the Figueroa Street Undercrossing Connector�Tseng and Penzien 1973�. Several assumptions were made tosimplify the structural model for applications to the Fast-MOSTexperiment. First, only the longitudinal response of the bridgestructure was considered. Second, axial deformations of the col-umn members were assumed negligible. Third, hinge connectionswere assumed between the column-deck interfaces. Consideringrotations and horizontal translations at each node, together withinternal moment releases, the structural model has 14 degrees offreedom. The redundancy of the substructures in the direction ofloading was found to be particularly useful in the developmentstages because if one remote site was not properly connected tothe master simulation, the simulation remained stable withoutcontribution to lateral force resistance from one column.
The selected bridge model provides for the extraction of fivesubstructures representing the columns, which can be modelednumerically or physically as specimens in the laboratory. Further,the column-deck pin connections simplify the experimental sub-structures to a single degree of freedom, that is, a cantilever sub-jected to a lateral force by a single pin-connected actuator. Fig. 4shows the name assigned to each column �COL01, COL02,COL03, COL04, and COL05� and location of each remote sub-structure �Berkeley, Boulder, UIUC, Buffalo, and Lehigh, respec-tively�. More details on the physical specimens at remote sites aregiven in the next section.
In the earthquake simulation, the bridge model was subjectedto 15 s of ground shaking using a synthetic ground motion. The
Fig. 4. Six-span bridge mode
numerical response of the structure was computed using the
JOUR
J. Struct. Eng. 2008.
operator-splitting integration algorithm �Hughes et al. 1979;Nakashima et al. 1990� combining implicit �for linear substruc-tures� and explicit �for nonlinear substructures� integration algo-rithms
Mai+1 + C�i+1 + KIdi+1 + Ki+1E di+1 = f i �1�
d̃i+1 = di + �t�i + ��t2/4�ai �2�
di+1 = d̃i + ��t2/4�ai+1 �3�
�i+1 = �i + ��t/2��ai + ai+1� �4�
One advantage of the operator-splitting algorithm is that it isbased on a predictor-corrector algorithm that does require itera-tion and offers unconditional stability for structures with stiffnesssoftening behavior. This algorithm was programmed as a moduleto the Java simulation coordinator to solve for the dynamic re-sponse of the bridge model subjected to seismic shaking. Anintegration time step of 0.01 s was used for a total of 1,500 simu-lation steps.
Remote Substructures
The purpose of Fast-MOST was to demonstrate NEESGrid capa-bilities for geographically distributed testing. To this end, the six-span bridge model was divided into numerical and physicalsubstructures and these substructures were distributed among theparticipating NEES equipment sites. Fig. 5 illustrates the distri-bution of the bridge column substructures to remote laboratoriesacross the United States. The simulation coordinator and the mas-ter numerical simulation including the deck model were located atBuffalo, N.Y.
Experiments were conducted between the Fall of 2003 and theFall of 2004. It should be noted that all NEES equipment sitesinvolved in Fast-MOST participated as remote sites. However,
eographically distributed tests
Fig. 5. Illustration of bridge model with geographically distributedsubstructures
not all sites were available during the scheduled simulations,largely due to scheduling conflicts, and at times participated withnumerical substructures. The results presented here correspond toa particular simulation conducted in September 2004 duringwhich University of California at Berkeley and University atBuffalo participated as experimental sites. These two laboratorieswere most accommodating because they had dedicated a small-scale testing setup for the development of hybrid testingalgorithms.
University of California at Berkeley Experimental Site
The experimental substructure at University of California at Ber-keley was located in the micro-NEES laboratory of thenees@berkeley equipment site. The model consists of a cantilevercolumn with an idealized plastic hinge connection at the base toperform tests in the nonlinear range. Fig. 6�a� shows a picture oftwo identical experimental setups; only one was used for the Fast-MOST. Based on preliminary testing of the cantilever, the initialstiffness of the test specimen was 0.49 kN /mm. with a peak re-sisting force of 8 kN. The yield displacement was approximately16 mm with maximum allowable displacements of 100 mm. Theforces and displacements were scaled to be representative of thefull-scale column substructure, where the scale factors were se-lected to match the initial stiffness of the specimen and the com-puted elastic stiffness of the full-scale numerical column model.The displacement scale factor was also selected to ensure that theexpected peak displacements remained within the capacity of thetesting system. These scale factors were selected such that for a
Fig. 6. Photograph of experimental substructures at: �a� Berkeley; �b�Buffalo used to physically model bridge columns
low-level linear simulation the same results are obtained from an
all-computational substructure simulation and a hybrid simulationwith numerical and experimental substructures.
University at Buffalo Experimental Site
Fig. 6�b� shows a similar cantilever-type specimen at nees@buffalo that was used to simulate the restoring forces correspond-ing to the column substructure at the University at Buffalo. Thereduced scale substructure had an initial stiffness of 0.32 kN /mm,yield displacement of 8 mm, and a displacement capacity of25 mm. The displacement and forces were also scaled to matchthe initial stiffness of the numerical full-scale column.
Testing Procedure
An incremental and modular testing procedure was employed totest operation of all the computer and control systems utilized inthe test:1. Each remote site initially ran a local simulation of all the
sites to ensure their software environment was properly con-figured to run the NTCP server and simulation coordinator;
2. The local simulation was repeated with the local experimentenabled as a substructure with the simulation to test theNEESGrid interface to the control system. Further, the simu-lation was conducted at low levels to avoid damage to thespecimens, or a computer model of the specimen was used athigh excitation levels. In either case the actual control hard-ware was used in the simulation;
3. The network connectivity and performance were verified byrunning a distributed test with all sites using numerical sub-structures; and
4. A final end-to-end test was performed to ensure the inte-grated system performed to specification. After verifying allcomponents, the actual hybrid simulation with experimentalsubstructures was carried out.
Analysis of Test Results
The results from the distributed hybrid simulations carried out onSeptember 9, 2004 are presented in this section. During thisexperiment, the master numerical simulation including the simu-lation coordinator was located at Buffalo. Two physical substruc-tures were included in the hybrid simulation: one at Berkeley andthe other at Buffalo. Lehigh participated in this experiment as aremote numerical substructure, while the column substructuresfor Boulder and University of Illinois at Urbana-Champaign weresimulated local to the master simulation in Buffalo. Numeroustrial simulations were conducted with all sites participating asexperimental and/or numerical substructures.
Simulation Results
The computed displacement response of the bridge model sub-jected to the selected ground acceleration record is presented inFig. 7. The figure shows the computed displacement at the5 degrees of freedom that connect to the remote substructures.These displacement histories can also be interpreted as the com-mand displacements generated for the five column substructures.Due to the axial rigidity of the bridge deck, the horizontal defor-mations of all the bridge columns along the longitudinal directionof the bridge are approximately the same. The small difference is
due to axial deformations of the bridge deck.
134:535-543.
Dow
nloa
ded
from
asc
elib
rary
.org
by
Uni
vers
ity o
f B
righ
ton
on 0
6/27
/14.
Cop
yrig
ht A
SCE
. For
per
sona
l use
onl
y; a
ll ri
ghts
res
erve
d.
Fig. 8 presents the force deformation behavior of the twoexperimental column substructures during the Fast-MOST simu-lation. Note that the nonlinear behavior demonstrated by thesubstructures from Berkeley and Buffalo are due to the actualmeasured behavior of the physical specimens. The hysteresiscurves are mainly the result of yielding of the experimental sub-structures, although they can also include stiction and other ac-tuator displacement control and force measurement errors. Thismeasured restoring force was considered in the numerical simu-lation of the bridge model. The remaining three column substruc-tures were simulated numerically.
Fig. 7. Computed displacement response of bridge model andcommand signal load history for experimental substructures �similarfor all substructures�
Fig. 9. Step duration of 1,500 steps recorded during dist
JOUR
J. Struct. Eng. 2008.
Step Timing and Network Delays
The Fast-MOST simulation lasted, 1074 s �about 18 min� for1,500 numerical integration steps. On average, the total step timeduration was 0.66 s. Fig. 9 presents detailed information on thetime duration per step. As shown in Fig. 9�a�, there were foursteps with substantial delays of approximately 23 s and one with3 s. These excessive delays were observed during other trial runswith a similar frequency of occurrence. The source of these de-lays is likely in the network communication, resulting from theloss of data packets that need to be resent after a timeout. In total,these network delays increased the total test time by 90 s. Moreimportantly, these long delays had a significant impact on theexperimental substructure behavior. Fig. 10 shows the measuredforce-displacement data of the Berkeley experimental substruc-ture between steps 1,050 and 1,100, which contains the third de-layed step. The drop in force as the displacement passes the 0 mmmark is caused by force relaxation as the actuator holds for sev-eral seconds during the delayed step. Based on this observation, itis important to minimize the occurrence of delays and achieve acontinuous load trajectory on the experimental substructures. Theevent-driven controller is intended to limit force relaxation errorto a few steps with severe delays as opposed to every single stepas in simulations using conventional ramp-and-hold loadingmethods.
Recovery from Network Failure
During the four steps with delays of over 20 s, the simulationcoordinator timedout and reissued the command for the current
step, which resulted in the simulation continuing, as opposed toending prematurely. This feature compensates for the unreliablenature of the internet and increases the reliability of the distrib-uted system and its ability to complete the simulation. Duringthese periods of network delay, the test was on hold with nomovement of the actuators, until network communications couldbe reestablished with all remote sites.
In other trial simulations, a more serious problem was encoun-tered that required one remote site to restart their simulation in themiddle of the experiment. To overcome this limitation, the simu-lation coordinator was configured to replay the experiment andbring the restarted site up to the current step by replaying thecommand for each step only to the failed site. Once the simula-tion coordinator reached the step in which the problem occurred,the simulation resumed normally with all the remote sites untilsuccessful completion of the experiment. This approach did notrequire the complete simulation to restart from the beginning.
Conclusion
A distributed control strategy was implemented into NEESGrid tosupport relatively fast and continuous hybrid simulation testingmethods with geographically distributed substructures takingabout 0.66 s per integration step. This performance is a substan-tial improvement compared to previously distributed tests whichrequired 13.2 s per step with more stringent security features. Thecontroller implemented for the faster strategy consists of a multi-threaded simulation coordinator to communicate with all the re-mote sites in parallel and an event-driven controller to generatecontinuous load histories for the experimental substructures thatminimize strain-rate and force-relaxation errors. These techniqueswere implemented in a simulation platform using secure commu-nication protocols developed for NEESGrid. The testing tech-niques developed here can be readily integrated into existinghybrid simulation platforms to enable geographically distributedtesting.
Acknowledgments
This work was supported in part by the NEES Program of theNational Science Foundation under Award Nos. CMS-0086621and CMS-0086611/12. Their support is gratefully acknowledged.
Fig. 10. Effects of time delays on yielded experimental substructureat Berkeley between steps 1,050 and 1,100.
Any opinions, findings, and conclusions or recommendations ex-
pressed in this material are those of the writers and do not neces-sarily reflect those of the National Science Foundation. The writ-ers worked closely with the NEES System Integration team �BillSpencer, Sridhar, Ghallapalli, and Laura Pearlman�, University ofColorado at Boulder �Benson Shing and Eric Stauffer�, LehighUniversity �Jim Ricles and Peter Bryant�, University of Illinois atUrbana Champaign �Narutoshi Nakata, Juan Carrion�, Universityof California at Berkeley �Andreas Shellenberg, Tony Yang�, andUniversity at Buffalo �Xiaoyun Shao, Scot Weinreber�. Their con-tribution is sincerely appreciated.
References
Campbell, S., and Stojadinovic, B. �1998�. “A system for simultaneouspseudodynamic testing of multiple substructures.” Proc., 6th U.S. Na-tional Conf. on Earthquake Engineering, Earthquake Engineering Re-search Institute, Seattle.
Chu, S. Y., Soong, T. T., and Reinhorn, A. M. �2005�. Active, hybrid andsemiactive structural control: A design and implementation handbook,Wiley, New York.
Harel, D. �1987�. “Statecharts: A visual formalism for complex systems.”Sci. Comput. Program., 8, 231–274.
Hughes, T. J. R., Pister, K. S., and Taylor, R. L. �1979�. “Implicit-explicitfinite elements in nonlinear transient analysis.” Comput. MethodsAppl. Mech. Eng., 17�18�, 159–182.
Magonette, G. �2001�. “Development and application of large-scale con-tinuous pseudodynamic testing techniques.” Philos. Trans. R. Soc.London, 359, 1771–1799.
Magonette, G., and Negro, P. �1998�. “Verification of the pseudodynamictest method.” Eur. Earthquake Eng., XII�I�, 40–50.
Mahin, S. A., Shing, P. B., Thewalt, C. R., and Hanson, R. D. �1989�.“Pseudodynamic test method—Current status and future direction.” J.Struct. Eng., 115�8�, 2113–2128.
Mosqueda, G., Stojadinovic, B., Hanley, J., Sivaselvan, M. V., and Rein-horn, A. M. �2005�. “Fast hybrid simulation with geographically dis-tributed substructures.” Proc., 1st Int. Conf. on Advances in Exp.Struct. Eng., AESE, Nagoya, Japan, 449–456.
Nakashima, M. �2001�. “Development, potential, and limitations of real-time online �pseudodynamic� testing.” Philos. Trans. R. Soc. London,359, 1851–1867.
Nakashima, M., Kaminosono, T., Ishida, I., and Ando, K. �1990�. “Inte-gration techniques for substructure pseudo dynamic test.” Proc., 4thU.S. National Conf. on Earthquake Engineering, Earthquake Engi-neering Research Institute, Palm Springs, Calif.
Nakashima, M., Kato, M., and Takaoka, E. �1992�. “Development ofreal-time pseudo dynamic testing.” Earthquake Eng. Struct. Dyn.,21�1�, 79–92.
Nakashima, M., and Masaoka, N. �1999�. “Real-time on-line test forMDOF systems.” Earthquake Eng. Struct. Dyn., 28�4�, 393–420.
Pan, P., Tada, M., and Nakashima, M. �2005�. “Online hybrid test byinternet linkage of distributed test-analysis domains.” EarthquakeEng. Struct. Dyn., 34�11�, 1407–1425.
Pearlman, L., D’Arcy, M., Johnson, E., Kesselman, C., and Plaszczak, P.�2003�. “NEESgrid teleoperation control protocol.” Technical Rep No.NEESgrid-2003–07, NEESgrid, �http://it.nees.org/� �Jan. 2008�.
Pearlman, L., D’Arcy, M., Plaszczak, P., and Kesselman, C. �2004�.“The NTCP control plugin.” Technical Rep. No. NEESgrid-2004-26,NEESgrid, �http://it.nees.org/�.
Spencer, B., et al. �2004a�. “The MOST experiment: Earthquakeengineering on the grid.” Technical Rep. No. NEESgrid-2004-26,NEESgrid, �http://it.nees.org/�.
Spencer, B., et al. �2004b�. “NEESGRID: A distributed collaboratory for
134:535-543.
Dow
nloa
ded
from
asc
elib
rary
.org
by
Uni
vers
ity o
f B
righ
ton
on 0
6/27
/14.
Cop
yrig
ht A
SCE
. For
per
sona
l use
onl
y; a
ll ri
ghts
res
erve
d.
advanced earthquake engineering experiment and simulation.” Proc.,13th World Conf. on Earthquake Engineering, Vancouver, Canada,Paper No. 1674.
Stojadinovic, B., Mosqueda, G., and Mahin, S. A. �2006�. “Event-drivencontrol system for geographically distributed hybrid simulation.” J.Struct. Eng., 132�1�, 68–77.
Takahashi, Y., and Fenves, G. L. �2006�. “Software framework for dis-tributed experimental-computational simulation of structural sys-tems.” Earthquake Eng. Struct. Dyn., 35�3�, 267–291.
Takanashi, K., and Nakashima, M. �1987�. “Japanese activities on on-linetesting.” J. Eng. Mech., 113�7�, 1014–1032.
Takanashi, K., Udagawa, K., Seki, M., Okada, T., and Tanaka, H. �1975�.“Nonlinear earthquake response analysis of structures by a computer-actuator on-line system �details of the system�.” Trans. of the Archi-tectural Institute of Japan, 229, 77–83.
P.-C. �2003�. “Seismic hazard mitigation: Internet-based hybrid test-ing framework and examples.” Proc., Int. Colloquium on Natural
Hazard Mitigation: Methods and Applications, France.Tseng, W. S., and Penzien, J. �1973�. “Analytical investigation of the
seismic response of long multiple span highway bridges.” Rep. No.UCB/EERC-73/12, Earthquake Engineering Research Center, Univ. ofCalifornia, Berkeley, Calif.
Wang, K. J., Lin, B. Z., Cheng, W. C., Wang, S. J., Yang, Y. S., and Tsai,K. C. �2005�. “A platform for networked collaborative structural ex-periments.” Proc., 1st Int. Conf. on Advanced in Experimental Struc-tural Engineering, Nagoya, Japan.
Watanabe, E., Kitada, T., Kunitomo, S., and Nagata, K. �2001�. “Parallelpseudodynamic seismic loading test on elevated bridge systemthrough the Internet.” Proc., 8th East Asia-Pacific Conf. on Structural
