Realistic Case Studies of Wireless Structural Control

Bo Li1, Zhuoxiong Sun2, Kirill Mechitov3, Gregory Hackmann1, Chenyang Lu1,Shirley J. Dyke2, Gul Agha3, Billie F. Spencer Jr.4

1Department of Computer Science and Engineering, Washington University in St. Louis2School of Mechanical Engineering, Purdue University

3Department of Computer Science, University of Illinois at Urbana-Champaign4Department of Civil and Environmental Engineering, University of Illinois at Urbana-Champaign

ABSTRACTWireless Structural Control (WSC) systems can play a crucial rolein protecting civil infrastructure in the event of earthquakes andother natural disasters. Such systems represent an exemplary classof cyber-physical systems that perform close-loop control usingwireless sensor networks. Existing WSC research usually employswireless sensors installed on small lab structures, which cannotcapture realistic delays and data loss in wireless sensor networksdeployed on large civil structures. The lack of realistic tools thatcapture both the cyber (wireless) and physical (structural) aspectsof WSC systems has been a hurdle for cyber-physical systems re-search for civil infrastructure. This advances the state of the artthrough the following contributions. First, we developed the Wire-less Cyber-Physical Simulator (WCPS), an integrated environmentthat combines realistic simulations of both wireless sensor networksand structures. WCPS integrates Simulink and TOSSIM, a state-of-the-art sensor network simulator featuring a realistic wirelessmodel seeded by signal traces. Second, we performed two real-istic case studies each combining a structural model with wire-less traces collected from real-world environments. The buildingstudy combines a benchmark building model and wireless tracescollected from a multi-story building. The bridge study combinesthe structural model of the Cape Girardeau bridge over the Mis-sissippi River and wireless traces collected from a similar bridge(the Jindo Bridge) in South Korea. These case studies shed lighton the challenges of WSC and the limitations of a traditional struc-tural control approach under realistic wireless conditions. Finally,we proposed a cyber-physical co-design approach to WSC that in-tegrates a novel holistic scheduling scheme (for sensing, communi-cation and control) and an Optimal Time Delay Controller (OTDC)that substantially improves structural control performance.

Categories and Subject DescriptorsC.2.4 [Distributed Systems]: Distributed applications; C.2.1 [NetworkArchitecture and Design]: Wireless communication

Permission to make digital or hard copies of all or part of this work forpersonal or classroom use is granted without fee provided that copies arenot made or distributed for profit or commercial advantage and that copiesbear this notice and the full citation on the first page. To copy otherwise, torepublish, to post on servers or to redistribute to lists, requires prior specificpermission and/or a fee.ICCPS’13 April 8–11, 2013, Philadelphia, PA, USA.Copyright 2013 ACM ACM 978-1-4503-1996-6/13/04 ...$15.00.

KeywordsWireless structural control, Cyber-physical system, Wireless sensornetwork, Wireless cyber-physical simulator

1. INTRODUCTIONWireless Structural Control (WSC) is a promising cyber-physical

system technology for protecting our civil infrastructure in the eventof earthquakes and other natural disasters. A WSC system employsa feedback control loop to control the dynamic response of a civilstructure based on sensor data collected through wireless sensornetworks. As a representative example of cyber-physical systems,a WSC system requires holistic system designs that crosscut cyber(wireless and control) and physical (structural dynamics) compo-nents.

Since deployments of control systems on large civil structures(e.g., bridges and buildings) are costly and labor intensive, to dateWSC systems have mostly been evaluated using wireless sensorsinstalled on small lab structures. Unfortunately, such networks can-not capture the delays and data loss in wireless sensor networks de-ployed on large civil structures in real-world environments. Thereis a critical need for simulation tools and case studies that realisti-cally model wireless characteristics and the structural dynamics ofWSC systems.

To meet this challenge in WSC research, we have developed asimulator specifically designed to support realistic simulations ofwireless cyber-physical systems. Specifically, the contributions ofthis paper are three-fold:

• First, we describe the Wireless Cyber-Physical Simulator(WCPS),an integrated environment that combines realistic simulationsof both wireless sensor networks and structures. WCPS in-tegrates Simulink and TOSSIM, a state-of-the-art sensor net-work simulator featuring a realistic wireless model seeded byreal-world wireless traces.

• Second, we present two realistic case studies each matchinga structural model with wireless traces collected from real-world environments. The building study combines a bench-mark building model and wireless traces collected from amulti-story building. The bridge study combines the struc-tural model of the Cape Girardeau Bridge over the Missis-sippi River and wireless traces collected from a similar bridge(the Jindo Bridge) in South Korea. These case studies shedlights on the challenges of WSC and the limitations of tradi-tional structural control approaches.

• Finally, we propose a cyber-physical co-design approach toWSC that integrates a holistic scheduling scheme (includingsensing, communication and control) and an Optimal Time

Delay Controller (OTDC), which substantially improves struc-tural control performance in the presence of wireless com-munication delay and packet loss.

While this paper focuses on WSC as case studies, the WCPStool can be used to simulate other wireless control systems. Fur-thermore, our cyber-physical co-design approach and insights fromthe case studies can be generalized to other cyber-physical systems,especially large-scale wireless control systems. WCPS has been re-leased as open-source software at http://wcps.cse.wustl.edu.

The rest of the paper is organized as follows. Section II reviewsrelated works. Section III presents the WCPS simulator. SectionIV describes explicit designs of the case studies. Section V detailsthe cyber-physical co-design approach to wireless control. SectionVI presents the results of the case studies. Section VII concludesthe paper.

2. RELATED WORKWireless Structural Health Monitoring(WSHM) research has been

active in the past decade [13, 16]. Recent efforts for WSHM in-clude: a distributed wireless sensing system for WSHM [21], thefirst wireless system deployed on a tower of over 600 meters tall [23], a networked computing approach in WSHM [15], a high qual-ity sensor placement study for WSHM [20], a cyber-physical co-design of wireless distributed structural health monitoring [12], thelargest wireless bridge monitoring system in the world [14] andTorre Aquila deployment for heritage building monitoring [6]. Toname a few.

However, close-loop wireless control for civil structures is stillin its infancy. While early efforts developed control algorithms andprototype wireless control systems [22, 32, 33], all the previous ex-periments were performed on small-scale lab structures. In the labsettings, wireless sensors within a single hop and experience nodata loss due to physical proximity of the devices.

Wireless control has been studied with promising results in otherdomains [3–5,24]. The challenge in realistic experimentation withWSC systems motivates the development of our WCPS and casestudies based on real-world wireless traces and realistic structuralmodels. Our work thus expands the field of wireless control to thecivil infrastructure domain. Moreover, WCPS can also be used tosimulate other large-scale wireless control systems, and our scheduling-control co-design approach may be generalized to other wirelesscontrol systems. Research works using WCPS yet with more focuson civil structural analysis are introduced in [30, 31]

Truetime [7] is a well established control system simulator thatenables holistic studies of CPU scheduling, communication andcontrol algorithms. While Truetime supports wirelss networks, itswireless models are relatively simple and do not capture complexproperties of wireless sensor networks such as probabilistic andbursty packet receptions and irregular radio properties [29]. Inaddition, Truetime implements wireless models within Simulink.While a native implementation may improve efficiency, it cannotleverage existing wireless simulators that implement sophisticatedwireless models.

NCSWT [11] is a recent simulator for wireless cyber-physicalsystems. Instead of implementing wireless simulations natively, itintegrates with the NS-2 simulator with support for wireless net-works. While WCPS shares a similar federated approach to in-corporate an existing wireless simulator, we choose to integrateWCPS with TOSSIM [19] which features a more realistic wirelesssensor network model than NS-2. Despite its wide adoption as anetwork simulator, the wireless models in NS-2 suffers from being

incapable of capturing the probabilistic and irregular packet recep-tions that are common in low-power wireless networks. Leveragingnoise traces and statistical models, TOSSIM can capture complextemporal link dynamics that are crucial for realistic cyber-physicalsystems modeling. As the standard TinyOS simulator, TOSSIM hasbeen widely used for wireless sensor network research and has beenvalidated in diverse real-world environments [18]. Moreover, thetrace-driven simulation approach of TOSSIM enables us to studythe impacts of different wireless environments. We also providethe first set of realistic case studies based on real-world wirelesstraces, as well as a novel scheduling-control co-design approach toWSC.

3. WCPSThis section describes the design and implementation of WCPS [1].

Actuator

Actuator

Basestation

Wireless node

Wireless node

Wireless node

Wire

connection

Wire

connection

Figure 1: WSC System Model.

WCPS supports a general wireless control system model shownin Fig. 1. A wireless control system consists of a set of wireless sen-sors, a controller and a set of actuators. The sensors form a wirelessmesh network connected with a base station hosting the controller.Since the controller is usually located close to the actuators in WSCsystems, we assume the base station and actuators are connected bya wired network whose latency is negligible compared to that in thewireless sensor network. Following the centralized network man-agement approach of the WirelessHART standard [26, 27], WCPSemploys a centralized network manager to compute routing andtransmission schedules for the wireless sensor network.

WCPS employs a federated architecture that integrates (1) Simulinkfor simulating the physical system (structural) dynamics and thecontroller and (2) TOSSIM for simulating the wireless sensor net-work. Simulink has been widely used by control and structural en-gineers to design and study structural control systems, while TOSSIMis specifically designed to simulate wireless sensor networks basedon realistic wireless link models that have been validated in di-verse real-world environments [18]. By combining Simulink andTOSSIM, WCPS provides an integrated environment to simulatewireless control systems in a holistic and realistic fashion.

As shown in the architecture illustrated in Fig. 2, WCPS sim-ulates the feedback control loop of the control system as follows.Sensor data is generated from structural models. Through a cross-platform function call from Simulink, sensor data is injected to thecorresponding wireless sensors in TOSSIM. Following the routesand transmission schedule calculated by the network manager mod-ule, TOSSIM simulates the end-to-end wireless communication ofthe sensor data packets from the sensors to the base station, andthen return the packet delay and loss to the Interfacing Block inSimulink through the Python interface. The Packet Collector mod-ule then to extracts packet delivery information(the delay and loss)

from the message pool of returned values in Simulink. Sensordata and their loss and delay are then provided to the Data Block,which then feed the sensor data to the controller at the right timebased on the packet delay (if the packet is not lost). WCPS uti-lizes basic API (e.g., the dos, UNIX command) of MATLAB todo cross-platform function calls. In TOSSIM, we re-implement aprintf method in TinyOS to send TOSSIM simulation results to theInterfacing Block.

Structural controller

Packet Collector

Message pool

Cross-platform function call

Python interface

Wireless signal

EarthquakesStructure model

Scheduling

Network Manager

Routing

Wireless noise

Routing layer

TDMA MAC layer

Wireless link model

Routing

Schedule

RSSI

Noise

Exctitaion

Sensor data

Controlinformation

Sensor data after delay and loss

Return values of function call TOSSIM

Simulink

User Inputs

Sensor data

Data Block

Interfacing Block

Data with SS/UDS decisions

Wireless Network

Figure 2: Component architecture of WCPS.

User inputs to WCPS includes excitation signals to the structure(e.g., acceleration caused by earthquakes) and wireless traces usedas input to TOSSIM. Excitation signal of the structure is providedto the structure models in the format of MAT files.

The scheduler module calculates transmission schedules. Net-working schedule is then deployed into the MAC layer code ofwireless nodes and becomes effective after a TinyOS compilation.The TDMA MAC layer in WCPS is developed based on the MACLayer Architecture (MLA) library [17] and further adapted forTOSSIM under TinyOS 2.1.1. Received Signal Strength Indica-tion(RSSI) and wireless noises traces are collected from real-worldenvironments and provided to the wireless model [18] used by TOSSIMfor realistic wireless network simulations.

As shown in Fig. 2, the interfaces between the Simulink modeland TOSSIM are encapsulated as two MATLAB embedded func-tions in Simulink: the Interfacing Block and the Data Block. TheInterfacing Block extracts delay and loss information from TOSSIMmessages, and the Data Block decides what data will be used fordiscrete control during each sampling period. The federated archi-tecture of WCPS provides great flexibilities to incorporate differ-ent structural models and implement alternative scheduling-controlapproaches. Further details about WCPS (including user manual,documentation and the source code) is available at http://wcps.cse.wustl.edu.

4. CASE STUDY DESIGNThis section presents the design of the case studies on wireless

control of a three-story building and a bridge, respectively.

4.1 Excitation SignalTo study structural response to an earthquake, we use measure-

ments from a real earthquake as the excitation signal in both casestudies. As shown in Fig. 3 the excitation signal was recorded atthe Imperial Valley Irrigation District substation in El Centro, Cal-ifornia, during the Imperial Valley, California earthquake of May

0 10 20 30 40 50−3

−2

−1

0

1

2

3

4

Time (Second)

Acce

lera

tio

n (

m/s

2)

El Centro Earthquake

Figure 3: The El Centro earthquake as excitation signal ofstructural control [10] [28].

18, 1940 [10]. The EI Centro earthquake last 50 seconds with amaximum acceleration of 3m/s2 at the beginning.

4.2 Design the Building Study

4.2.1 Wireless trace collectionThe wireless sensor network in this study comprises a base sta-

tion and four distributed sensors. Wireless traces were collectedfrom Bryan Hall of Washington University. The base station is lo-cated on floor 3, and the wireless sensors (TelosB motes [25]) areplaced on floor 0, 1, 2, and 3, respectively. The sensors recordRSSI and noise traces on channel 26 of the IEEE 802.15.4 radio.Each TelosB mote is equipped with a Chipcon CC2420 radio withits transmission power set to 0 dBm. Our measurements show thatthe wireless signal of the TelosB motes can go through at most twofloors. As a result the sensor on floor 0 needs a multi-hop route tosend data to the base station on floor 3.

4.2.2 Building modelOur building model is based on a three-story test structure shown

in Fig. 5(a) [28]. The test structure is subject to one-dimensionalground motion. The frame is constructed of steel, with a heightof 158 cm. For control purposes, a simple implementation of anActive Mass Driver (AMD) is placed on the 3rd floor of the teststructure (see Fig. 5(b)). The AMD actuation system has a sin-gle hydraulic actuator with steel masses attached to the ends of thepiston rod. Since hydraulic actuators are inherently open-loop un-stable, position feedback is employed to stabilize the actuator. Theposition of the actuator is obtained from an LVDT (linear variabledifferential transformer), rigidly mounted between the end of thepiston rod and the third floor. The first three modes of the test struc-ture are 5.81 Hz, 17.68 Hz and 28.53 Hz, with associated dampingratios of 0.33%, 0.23%, and 0.30%, respectively [28].

We developed a Simulink model (shown in Fig. 4) with refer-ence to the steel test structure at a 1:1 ratio. The Simulink model isdesigned to simulate a real-world three-story building, with map-ping ratios of: force = 1:60, mass = 1:206, displacement = 4:29and acceleration = 7:2, and time =1:5. Since the time scales of theSimulink model and a real-world building have a 1:5 ratio, the natu-ral frequencies of the model are approximately five times as large as

YmeaYout

Z Vector toWorkspace

ZOUT_in_r

Ydelay

Unit Delay

1/z

Time Vector to Workspace

TIME

SensorNoises

Saturation SaturationQuantizer2Quantizer1

Mux

Mux

Interfacing Block

ymea

ymea1

ydelay

Tossim_block

Earthquake

ELC_A.MAT

Discrete Controller

y(n)=Cx(n)+Du(n)x(n+1)=Ax(n)+Bu(n)

Demux

Data Block

yout

ydelay

yout_delay delay_block

Command Signal to Workspace

usignal

Clock

Building Model

x’ = Ax+Bu y = Cx+Du

Figure 4: Simulink diagram for wireless building control.

(a)

(b)

Figure 5: Building control system [28]: (a) The 3-story teststructure; (b) Active Mass Driver actuation system.

those of a real-world building. Previous Simulink implementationsof the building model was modeled as a continuous system and atime step of 0.0001 second was used to reduce integration errors. InWCPS we further discretize the Simulink model and perform step-by-step simulations with a step length of 1 ms, which correspondedto 5 ms in a real-world building. As the network used 10 ms slotsfor TDMA, a slot in the simulated wireless network therefore cor-respond to two run steps of the Simulink model.

As shown in Fig. 4, the structural response signal is first gener-ated by the building model, then converted by the Analog to DigitalConverter (ADC) to digital values, and fed to TOSSIM. TOSSIMdelivers the sensor data along with its status (loss and delay) to the

discrete controller. The output of the controller is then convertedfrom digital values to analog signals by the Digital to Analog Con-verter (DAC). Eventually, sensor data with control information arefed back to the building model, which closes the control loop.

4.3 Design the Bridge StudyThe bridge study simulates wireless control of the Cape Girardeau

bridge in Missouri, USA. The cable-stayed bridge (see Fig. 9(b)) isthe Missouri 74 Illinois 146 bridge spanning the Mississippi Rivernear Cape Girardeau, Missouri, designed by the HNTB Corpora-tion. Since no wireless sensors have been deployed on the bridge,we opt to use wireless traces collected from a wireless sensor net-work deployed on the Jindo bridge [14], South Korea, which sharessimilar dimensions (e.g., tower height and span range) and designswith the Cape Girardeau bridge. The sensor placement of the Jindodeployment is then mirror mapped onto the Cape Girardeau bridge.This approach takes advantage of the flexibility of WCPS to com-bine structural models and wireless traces from different (but simi-lar) structures for integrated WSC simulations.

Pylon

Sensor

Base

Station

Figure 6: Wireless pylon sensor and base station placement onthe Jindo bridge.

First Generation Benchmark Control Problem for Cable−Stayed Bridges

S.J. Dyke, G. Turan, J.M. Caicedo, L.A. Bergan, and S. Hague

yout_d1

Ydelay_d

yout_d

Yout_d

yin_d

Yin_inte

discrete state estimator

y(n)=Cx(n)+Du(n)x(n+1)=Ax(n)+Bu(n)

To Workspace1

ysensor

To Workspace

Ysen_inte

Time

TIME

Sensors

ymys

Saturation1

Saturation

Quantizer1

Quantizer

Mux

MultiportSwitch1

1

2

*, 3

Interfacing Block

ymea

ymea1

ydelay

Tossim_block

Integer Delay1

−1Z

Gain

K*u

mexico.mat

band_noise.mat

elcentro.matEvaluation

Outputs

Ye

Discrete State−Space

y(n)=Cx(n)+Du(n)x(n+1)=Ax(n)+Bu(n)

Device Outputs

Yf

Data Block

yout

ydelay

yout_delay delay_block

Control Devices

u

f

yf

Eqnum

Clock

Bridge Model

Ground Excitation

Sim Outputs

ye

yc

ym

Figure 7: Simulink diagram for wireless bridge control.

4.3.1 Wireless trace collectionThe Jindo deployment utilizes the MEMSIC Imote2 platform

and a total of 113 Imote2 sensor nodes with 659 distinct sensorchannels. Each node integrates the Imote2, the ISM400 sensorboard, and a rechargeable battery supplied by a solar panel. Com-bined with the Illinois SHM Services Toolsuite [2], these power-ful nodes allow for synchronized data collection, aggregation, syn-thesis and decision-making in real time. The system has success-fully captured ambient traffic loading with peak acceleration rang-ing from less than 5 mg to over 30 mg. Further analysis of the dataresulted in the successful identification of the first twelve modes ofvibration on the deck, as well as tension forces of 10 cables withlarge tensile stresses [14]. To serve as input to the TOSSIM sim-ulation, a subset of Imote2 nodes located along deck of the bridgeand sensors on the top of the pylons are selected for wireless tracecollection. With wireless traces collected from the Jindo bridge,we are able to build a 58-node routing network in TOSSIM for theCape Girardeau bridge.

During our wireless trace collection on the Jindo bridge, sen-sors located on the top of the pylons pose a special case for tracecollection. Whereas the sensors on the bridge deck form a con-nected graph, the pylon nodes are isolated. Due to the height ofthe pylons, these nodes are outside the maximum radio range ofthe deck nodes. In our Jindo deployment, pylon sensors are fittedwith directional antennas, which are pointed away from the bridgedeck towards a base station, located on the nearby Jindo Bridge(seeFig. 6). For the purpose of modeling a connected network, the reallink quality measurements between the pylon sensor and base sta-tion node are mapped onto a virtual link in TOSSIM. During thenetwork mapping, as the distances involved in Jindo bridge andCape Girardeau bridge are similar and both bridges are in open ar-eas, we assume this network mapping would correspond closely toa real wireless network setup.

Based on the structural model [10] we select sensor 240 and 353located on the tow towers of the Cape Girardeau bridge, sensors151 and 185 at the foots of towers and sensor 34 in the mid-spanfor structural control. Acceleration and displacement readings fromthe five selected sensors are sent to the base station located nearsensor 185 using routes with the minimum ETX in the network. Totest the accuracy of the TOSSIM simulation, we implement a testapplication in TOSSIM and compare the Packet Reception Ratio(PRR) of the simulation with that from the field test in Jindo. Fig. 8plots the Cumulative Distribution Function (CDF) of the PRR dif-ference between the field measurements and the TOSSIM simula-

0 20 40 60 80 1000

0.2

0.4

0.6

0.8

1

Difference of PRR (%)

CD

F

Figure 8: PRR Difference between field measurement andTOSSIM simulation.

tions for all 467 wireless links. Of all the wireless links, over 85%of them have the same PRR in the field measurements and the simu-lation, indicating TOSSIM can deliver high fidelity link simulationsbased on real-world traces.

4.3.2 Bridge modelA high-fidelity Cape Girardeau bridge model (see Fig. 9(b)) was

incorporated in WCPS for bridge control. A linear evaluation modelwas used for evaluation of the benchmark bridge model. How-ever, the stiffness matrices used in this linear model are those ofthe structure determined through a nonlinear static analysis cor-responding to the deformed state of the bridge with dead loads.Experimental study indicates that the longitudinal direction of thebridge is most destructive [10].

For control purposes the joints between the tower and the deckare disconnected and replaced by the control devices. As expected,the frequencies of this model are much lower than those of the nom-inal bridge model after incorporating the control device. The firstten frequencies of this second model are 0.1618, 0.2666, 0.3723,04545, 0.5015, 0.5650, 0.6187, 0.6486, 0.6965, and 0.7094 Hz [10].

Fig. 7 shows the block diagram of the wireless bridge controlsystem. Similar to the building control, the structural response ofthe bridge will go through ADC, a wireless network simulated inTOSSIM, and a discrete state estimator. The control inputs are con-verted by DAC to analog signals sent to the actuator.

(a)

(b)

Figure 9: Cape Girardeau model in WCPS: (a) the Cape Gi-rardeau bridge; (b) Simulink model of the Cape Girardeaubridge [10].

5. WIRELESS CONTROL APPROACHESWe implement and compare two alternative control approaches

to WSC. Instead of isolating the designs of the control algorithmand wireless sensor networks, we study holistic cyber-physical co-designs that integrate control algorithms and scheduling strategiesfor data collection, communication and utilization. As a baselinedesign the first approach integrates a traditional structural controlalgorithm called the Sample Controller (SC) [28] and a schedulingstrategy that minimizes sensing delays. The second approach inte-grates the Optimal Time Delay Controller (OTDC) [9] and a novelscheduling strategy that lead to uniform sensing delays. Note thatboth SC and OTDC controllers were originally designed for wiredstructural control. Our work provides the first case studies of thesecontrol algorithms when applied to wireless structural control.

5.1 Sample ControllerSC employs the Linear Quadratic Gaussian (LQG) optimal con-

trol algorithm [28]. LQG is a combination of linear quadratic esti-mator (LQE) and linear quadratic regulator (LQR). The cost func-tion to be minimized in SC is defined in Equation 1, where xr is thereduced states vector, u is the control force,CZr andDZ

r are systemmatrices for the regulated output vector, and Q and R are weight-ing matrices. More details of SCc can be found in [28] and [10].

J = limτ→∞

1

τE

[∫ τ

0

{(CZr x

r +DZr u)T

Q(CZr x

r +DZr u)

+uTRu}(dt)

](1)

Specifically for SC, we implement a Sequential Scheduler (SS)which schedule one packet each TDMA time slot. Key data utiliza-tion mechanism for SS is to transmit the latest available data. Forexample, given vector [yx1 , yx2 , yx3 , yx4 ] as the data collected by sen-sor 1, 2, 3, 4 at the beginning of slot x, sensor 3 at the beginning of

ControlSensing

DelayData eligible for control use

Sensor1

1 2 3

Sensor 2

Sensor 3

Sensor 4

4->0

3->0

2->0

1->33->0

4 5 6 7 8 9 104->0

3->0

2->0

1->33->0

11

Figure 10: Example of baseline SC controller with SequentialScheduler.

slot 2 (see Fig. 10) chooses to transmit y23 instead of y13 because y23is the latest reading. Similarly, sensor 2 chooses to transmit y32 atthe starting point of slot 3 because y32 is the most up-to-date read-ing. SS makes sure that only latest sensor data is used for control,but it also sacrifices sensing synchronizations.

Fig. 10 illustrates working mechanism of SC and SS with a four-sensor network example. Sensor 1, 2, 3, 4 are located on floor 1, 2,3, 4 of a building while the base station is located on the 4th floor.Since sensor 4, 3, and 2 have 1-hop distance to the base station,each needs one time slot to its data to the base station, while sensor1 needs two because it is two hop away from the base station. SCcontrol (denoted by dark arrows in Fig. 10) starts at the end of slot1 with data vector [0, 0, 0, y14 ] as only the first reading (collected inslot 1) of sensor 4 has arrived. By the end of the slot 2, SC com-putes its control input with [0, 0, y23 , y

14 ] because the second reading

(collected in slot 2) of sensor 3 has arrived. By the end of slot 3, SCuses [0, y32 , y23 , y14 ] to compute its control input. The same data vec-tor is used again at the end of slot 4 because no reading from sensor1 has arrived yet. By the end of slot 5, SC uses [y41 , y32 , y23 , y14 ] forcontrol, which completes a communication cycle from all sensors.Starting from slot 6, another cycle of data collection and controloccur using the same schedule. Intuitively, the combination of SCand SS aims to reduce the delay of the sensor data used for control.Henceforth, we refer to the control scheme combining SC and SSas SC for simplicity.

5.2 Optimal Time Delay ControllerOTDC [9] was originally designed for constant-delay system as

shown in Equation 2, where l is the time delay. OTDC is designedto minimize the cost function J by selecting an optimal controlforce pd in Equation 3. However, in a wireless sensor networkdata from different sensors will be delivered to the controller at dif-ferent delays. To use OTDC to WSC effectively we design a novelscheduling strategy called the Uniform Delay Scheduler (UDS) thatpushes sensor data to the controller at uniform delays.

z [k + 1] = Az [k] +Bpd [k − l] (2)

Fig. 11 illustrates the schedule produced by UDS for the samefour-sensor example used in the last subsection. UDS first buffersone batch of data (five readings for each sensor). Afterwards, acycle of five time slots is used to deliver the batched data to the

ControlSensing

DelayData eligible for control use

Sensor1

1 2 3

Sensor 2

Sensor 3

Sensor 4

4->0

3->0

2->0

1->3 3->0

4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 204->0

3->0

2->0

1->3 3->0

Figure 11: Example of OTDC-1 with UDS scheduler.

base station. By the end of slot 10, OTDC starts with data vector[y11 , y

12 , y

13 , y

14 ], followed by [y21 , y

22 , y

23 , y

24 ] in the next time slot,

and [y31 , y32 , y

33 , y

34 ] in the time slot after. This pattern continues

till the end of slot 14. In time slot 15 OTDC starts a new cycleand uses [y61 , y

62 , y

63 , y

64 ] by the end of time slot 15. Under UDS

data from different sensors shares a uniform network delay (10 timeslots, or 100 ms in Fig. 11). UDS therefore trades one cycle ofdelay for uniform delays among sensors. This feature makes UDSparticularly suitable for OTDC specifically designed for systemswith constant delays. As shown in our case studies this scheduling-control co-design approach leads to an effective WSC system.

J |pd =

∞∑k=l

(zTd [k]Qzd (k) + pd

T [k − l]Rpd [k − l])

(3)

ControlSensing

DelayData eligible for control use

Sensor1

1 2 3

Sensor 2

Sensor 3

Sensor 4

4->0

3->0

2->0

1->33->0

4 5 6 7 8 9 104->0

3->0

2->0

1->33->0

11 12 13 14 15 16 17 18 19 20 21 22 23 24 254->0

3->0

2->0

1->33->0

4->0

3->0

2->0

1->33->0

Figure 12: Example of OTDC-2 with UDS scheduler.

Another challenge introduced by wireless networks is packet loss.Since the basic version of UDS described above schedules onlyone transmission attempt for each sensor reading, a packet dropmeans losing one batch of readings (e.g., 5 readings in Fig. 11). Todeal with packet loss we extend UDS to support multiple transmis-sions per sensor reading. Henceforth we use OTDC-k to denotea design that integrates OTDC and UDS that transmit each sensorreading k times. Due to the limited bandwidth of wireless sen-sor networks, OTDC-k retransmit sensor data from earlier cyclesby merging them into packets of later cycles. The simple packet-merging mechanism in OTDC-2 avoids costly retransmissions ofentire packets (e.g., as in WirlessHART). The number of batchesthat can be merged into a packet merging is limited by the packetpayload size, e.g., over 100 bytes for IEEE 802.15.4 packets.

For example, in Fig. 12, though the batch of data [y11 , y21 , y31 , y41 , y51 ]from sensor 1 may be available by the end of slot 10, OTDC-2 waitsfor one more cycle (five time slots) before pushing the sensor datato the controller. At the same time [y11 , y

21 , y

31 , y

41 , y

51 ] is merged

with [y61 , y71 , y

81 , y

91 , y

101 ] and goes through another cycle of net-

work communication. In this way, [y11 , y21 , y31 , y41 , y51 ] are transmit-ted twice and thus has better chance to be successfully delivered.OTDC-2 therefore trades additional network delay for higher relia-bility, while maintaining uniform delays across sensors. Increasingk in OTDC-k increases network delays while achieving higher re-liability.

6. RESULTS OF CASE STUDIESThis section presents the results of the case studies under real-

istic structural and wireless models in WCPS. In both case stud-ies we compare the performance of alternative wireless control ap-proaches, SC and OTDC-1. We also study the tradeoff betweendelay and data loss by comparing OTDC with different numbers ofretransmissions (OTDC-1, OTDC-2 and OTDC-3).

6.1 Wireless Building ControlThe building remains stable under all control approaches through-

out this case study. To evaluate the control performance we usethree categories of metrics: resource requirement, structural re-sponse, and constraints of the control system. We refer interestedreaders to [28] for detailed definitions of these metrics. We per-form simulations using four different wireless control approaches(SC, OTDC-1, OTDC-2, and OTDC-3). Experimental results pre-sented below are from 25 simulations for each control approachand each simulation lasts 10,000 control steps.

Fig. 13 shows the end-to-end packet delivery ratio of the wire-less network. The end-to-end delivery ratio means the fraction ofpackets from the sensors that are successfully delivered to the con-troller. As shown in Fig. 13 Sensor 1 has the lowest delivery ratiobecause it has a 2-hop route to the controller. Recall that OTDC-1does not perform any retransmission, while OTDC-2 and OTDC-3performs retransmit each packet once and twice, respectively. Un-der OTDC-1 Sensors 1 and 4 have delivery ratios of 70% and over95%, respectively. As expected more retransmissions improve thedeliver ratios of all sensors at the cost of longer delays as describedearlier.

Fig. 14 shows the resource requirement of different control ap-proaches. OTDC-k approaches (see Fig. 14(a)) consistently re-quire less control power than SC. As k increases, OTDC-k re-quires slightly less control power. Similarly, as shown in Fig. 14(b),OTDC-1 reduces control force by 80% when compared to SC. Thedifferences in control force among different OTDC-k approachesare negligible. The results that OTDC-1 outperforms SC in bothmetrics indicate resource requirements are more sensitive to datasynchronization than to sensing delays in this building control sys-tem. OTDC-k with larger k results in negligible reduction of con-trol power and force, indicating resource requirements are not sen-sitive to network reliability in this case study.

The control performance regarding structural response is shownin Fig. 15. In term of peak inter-story drift in Fig. 15(a), OTDC-k achieves more reduction in inter-story drift than SC. Interest-ingly, higher k in OTDC-k increases peak inter-story drift. Re-call a higher k leads to higher communication reliability but longersensing delay. Inter-story drift is thus more sensitive to sensingdelays than to data loss in this case study. Similarly, as shown inFig. 15(b), OTDC-3 causes worse peak acceleration than all theother approaches. Hence, building structural responses are moresensitive to sensing delays than to data loss. In addition, OTDC-1 only slightly outperforms SC, which indicates limited impact ofdata synchronization on structural response.

The control performance regarding control system constraints isshown in Fig. 16. Fig. 16(a) and (b) plot the actuator peak ac-

OTDC−1 OTDC−2 OTDC−3

0.2

0.4

0.6

0.7

0.8

0.9

1

De

live

ry R

atio

Sensor 1

OTDC−1 OTDC−2 OTDC−3

0.2

0.4

0.6

0.7

0.8

0.9

1

De

live

ry R

atio

Sensor 2

OTDC−1 OTDC−2 OTDC−3

0.2

0.4

0.6

0.7

0.8

0.9

1

De

live

ry R

atio

Sensor 3

OTDC−1 OTDC−2 OTDC−3

0.2

0.4

0.6

0.7

0.8

0.9

1

De

live

ry R

atio

Sensor4

Figure 13: End-to-End Packet Delivery Ratio of Sensors in Building Study

SC OTDC−1 OTDC−2 OTDC−30

0.2

0.4

0.6

0.8

1

1.2

Re

qu

ire

d C

on

tro

l P

ow

er

(a)

SC OTDC−1 OTDC−2 OTDC−30

0.2

0.4

0.6

0.8

1

Re

qu

ire

d F

orc

e M

ag

nitu

de

(b)

Figure 14: Required Resource for Wireless Building Control: (a)Required Control Power; (b) Required Force Magnitude.

SC OTDC−1 OTDC−2 OTDC−30

0.2

0.4

0.6

0.8

1

Pe

ak I

nte

r−sto

ry D

rift

(a)

SC OTDC−1 OTDC−2 OTDC−30

0.2

0.4

0.6

0.8

1

Pe

ak A

cce

lera

tio

n

(b)

Figure 15: Structural Response under Wireless Building Control:(a) Peak Inter-story Drift; (b) Peak Acceleration.

celeration and Root Mean Square(RMS) acceleration, respectively.On both metrics OTDC-k approaches result in smaller actuator ac-celerations than SC. As k increases, we can see gradual decreasesin peak and RMS accelerations, indicating that these metrics aremore sensitive to improvement of communication reliability than tolonger sensing delays. In addition, the comparison between OTDC-1 and SC shows that the better data synchronization under OTDC-1has a larger impact than sensing delays.

In summary, we observe complex tradeoffs among data synchro-nization, sensing delay and communication reliability in wirelessbuilding control. Overall the OTDC approach combining a constant-delay control design and a scheduling scheme achieving data syn-chronization outperforms the SC approach that minimizes sensingdelay without data synchronization. This result highlights the effi-cacy of our control-scheduling co-design approach to wireless con-trol. Moreover, the design of the wireless communication protocolinvolves tradeoff between communication delay and data loss, witheach having stronger influence on different performance metrics.For our specific building study OTDC-1 and OTDC-2 outperformsOTDC-3. The complex tradeoff among multiple design aspectsconfirms the importance of a realistic simulation tool in designingwireless control systems.

6.2 Wireless Bridge ControlGiven the similarities in both the structural and wireless charac-

teristics shared by the Jindo bridge and the Cape Girardeau bridge,wireless traces collected from the Jindo bridge were used to simu-late the wireless sensor network used to control the Cape Girardeaubridge. The longest routing path is 3-hop. The results presentedbelow are from 25 simulations for each control case and each sim-ulation lasts 10,000 control steps. To mitigate large delays causedby large amount of packet deliveries for multiple sensors, networkscheduling in bridge control adopts an in-network aggregation ap-proach [8] through packet merging.

SC OTDC−1 OTDC−2 OTDC−30

1

2

3

4

4.5

Actu

ato

r P

eak A

ccele

ration

(a)

SC OTDC−1 OTDC−2 OTDC−30

1

2

3

Actu

ato

r R

MS

Accele

ration

(b)

Figure 16: System Constraint under Wireless Building Con-trol: (a) Actuator Peak Acceleration; (b) Actuator RMS Accel-eration.

The bridge network is highly reliable (99% PRR for almost alllinks with the Jindo trace) due to the relatively clean wireless en-vironment on the Jindo bridge as well as the fact that the Jindo de-ployment has line-of-sight sensor placement and strong radio an-tennas. As such, retransmission is not needed to achieve reliablecommunication. Therefore we only present the results of SC andOTDC-1 in this case study.

Since buildings and bridges have distinct structural properties,we adopt three different sets of metrics for performance evaluation.The metrics include maximum shear force, normalized shear forceand required control power. We refer interested readers to [10] forthe mathematical details of the metrics.

Fig. 17 plots the maximum shear force at the tower and the deckof the bridge. A smaller shear force is desirable in structural con-trol. SC performs slightly better in reducing the maximum towershear while OTDC-1 performs slightly better for reducing the max-imum deck shear, respectively. Fig. 18 plots the normalized shearforce at the tower and the deck. OTDC-1 slighly outperforms SCfor reducing normalized shear force.

SC OTDC−10

0.2

0.4

0.6

0.8

1

(a)

Maxim

um

Tow

er

Shear

SC OTDC−10

0.5

1

1.5

(b)

Maxim

um

Deck S

hear

Figure 17: Maximum Shear Force in Wireless Bridge Control: (a)Maximum Tower Shear; (b) Maximum Deck Shear.

SC OTDC−10

0.2

0.4

0.6

0.8

1

(a)

No

rma

lize

d T

ow

er

Sh

ea

r

SC OTDC−10

0.5

1

1.5

2

2.5

3

(b)

No

rma

lize

d D

eck S

he

ar

Figure 18: Normalized Shear Force in Wireless Bridge Control:(a) Normalized Tower Shear; (b) Normalized Deck Shear.

SC OTDC−10

0.5

1

1.5

2

2.5

3

3.5

4x 10

−3

(a)

Ma

xim

um

Co

ntr

ol P

ow

er

SC OTDC−10

1

2

3

4

5

x 10−4

(b)

To

tal C

on

tro

l P

ow

er

Figure 19: Control Power Requirement Performances forWireless Bridge Control: (a) Maximum Control Power; (b) To-tal Control Power.

While OTDC-1 did not show significant advantage over SC interm of shear force, it reduces both the required maximum controlpower and the total power requirement by nearly 50% comparedto SC (see Fig. 19). This result again demonstrated the effective-ness of the control-scheduling co-design approach adopted by theOTDC design.

7. CONCLUSIONWireless Structural Control (WSC) systems are a representative

class of cyber-physical systems that have the promise to protectour civil infrastructure in the event of earthquake and other naturaldisasters. To develop WSC systems it is critical to capture boththe cyber aspects (wireless communication and control) and thephysical aspects (structural dynamics) through realistic and holis-tic simulations. We have developed the Wireless Cyber-PhysicalSimulator (WCPS) that integrates a high-fidelity wireless simula-tor (TOSSIM) and a standard control system simulator (Simulink).With WCPS, we performed two case studies on structural controlsystems. Each case study combines a realistic structural model andwireless simulations driven by traces collected from real-world de-ployments. Our case studies leads to three important insights. First,there exist complex tradeoffs among data synchronization, sens-ing delay, and network reliability under realistic wireless structuralcontrol settings. Second, a realistic, integrated wireless controlsimulator like WCPS is critical in exploring the design tradeoffsin wireless control design. Finally, a control-scheduling co-designapproach is effective in wireless control design. In both case stud-ies the integration of a contant-delay control design and a schedul-ing scheme achieving data synchronization lead to substantial im-provement in control performance when compared to a traditionalcontrol design. Our cyber-physical simulation methodology andscheduling-control co-design approaches presented in this work not

only represent a promising step toward smart civil infrastructure,but also provide useful insights and tools that can be generalizedto other cyber-physical systems employing wireless control. TheWCPS tool and the case studies have been released as open sourcesoftware at http://wcps.cse.wustl.edu.

8. ACKNOWLEDGEMENTThis work was supported by NSF through CPS grants CNS-

1035773, CNS-1035748 and CNS-1035562, and NeTS grant CNS-1144552.

9. REFERENCES[1] http://wcps.cse.wustl.edu.[2] http://shm.cs.uiuc.edu.[3] J. Araujo, A. Anta, M. Mazo, J. Faria, A. Hernandez,

P. Tabuada, and K. Johansson. Self-triggered control overwireless sensor and actuator networks. In DCOSS, 2011.

[4] J. Bai, E. P. Eyisi, F. Qiu, Y. Xue, and X. D. Koutsoukos.Optimal cross-layer design of sampling rate adaptation andnetwork scheduling for wireless networked control systems.In ICCPS, 2012.

[5] G. Baliga, S. Graham, L. Sha, and P. Kumar. Etherware:Domainware for wireless control networks. In Proceedingsof The 7th IEEE International Symposium onObject-oriented Real-time Distributed Computing, 2004.

[6] M. Ceriotti, G. P. Picco, M. L. Amy, S. Guna, M. CorrL,M. Pozzi, D. Zonta, and P. Zanon. Monitoring heritagebuildings with wireless sensor networks: The torre aquiladeployment. In Proceedings of the 8th ACM/IEEEInternational Conference on Information Processing inSensor Networks (IPSN/SPOTS), San Francisco, CA, USA.

[7] A. Cervin, D. Henriksson, B. Lincoln, J. Eker, and K.-E.ArzOn. How does control timing affect performance?analysis and simulation of timing using jitterbug andtruetime. Proceedings of PWC 2003: Personal WirelessCommunication, Lecture Notes in Computer Science,23(3):16 – 30, June 2003.

[8] O. Chipara, C. Lu, J. Stankovic, and G.-C. Roman. Dynamicconflict-free transmission scheduling for sensor networkqueries. IEEE Transactions on Mobile Computing,10(5):734–748, May 2011.

[9] L. Chung, C. Lin, and K. Lu. Time-delay control ofstructures. Earthquake Engineering and StructuralDynamics, 24(5):687–701, 1995.

[10] S. Dyke, J. Caicedo, G. Turan, L. Bergman, and S. Hague.Phase i benchmark control problem for seismic response ofcable-stayed bridges. Journal of Structural Engineering,129(7):857–872, 2003.

[11] E. Eyisi, J. Bai, D. Riley, J. Weng, Y. Wei, Y. Xue, X. D.Koutsoukos, and J. Sztipanovits. Ncswt: An integratedmodeling and simulation tool for networked control systems.In The 15th International Conference on Hybrid Systems:Computation and Control (HSCC), 2012.

[12] G. Hackmann, W. Guo, G. Yan, C. Lu, and S. Dyke.Cyber-physical codesign of distributed structural healthmonitoring with wireless sensor networks. In ICCPS, pages119–128, April 2010.

[13] G. Hackmann, F. Sun, N. Castaneda, C. Lu, and S. Dyke. Aholistic approach to decentralized structural damagelocalization using wireless sensor networks. In RTSS, pages35–46, December 2008.

[14] S. Jang, H. Jo, S. Cho, K. Mechitov, J. Rice, S.-H. Sim, H.-J.Jung, C.-B. Yun, B. Spencer, and G. Agha. Structural healthmonitoring of a cable-stayed bridge using smart sensortechnology: deployment and evaluation. Smart Structuresand Systems, 6(5):439–460, 2010.

[15] A. Jindal and M. Liu. Networked computing in wirelesssensor networks for structural health monitoring. In SPIESymposium on Smart Structures and Materials,Nondestructive Evaluations and Health Monitoring, SanDiego, CA, March 2011.

[16] S. Kim, S. Pakzad, D. Culler, J. Demmel, G. Fenves,S. Glaser, and M. Turon. Health monitoring of civilinfrastructures using wireless sensor networks. In IPSN,2007.

[17] K. Klues, G. Hackmann, O. Chipara, and C. Lu. Acomponent-based architecture for power-efficient mediaaccess control in wireless sensor networks. In Sensys, 2007.

[18] H. Lee, A. Cerpa, and P. Levis. Improving wirelesssimulation through noise modeling. In IPSN, 2007.

[19] P. Levis, N. Lee, M. Welsh, and D. Culler. Tossim: Accurateand scalable simulation of entire tinyos applications. InSensys, 2003.

[20] B. Li, D. Wang, F. Wang, and Y. Ni. High quality sensorplacement for structural health monitoring systems:Refocusing on application demands. In Proc. IEEEINFOCOM’10, San Diego, CA, Mar., 2010.

[21] X. Liu, J. Cao, W.-Z. Song, and S. Tang. Distributed sensingfor high quality structural health monitoring using wirelesssensor networks. In The 33rd IEEE Real-Time SystemsSymposium (RTSS’12), 2012.

[22] J. Lynch, Y. Wang, R. Swartz, K.-C. Lu, and C.-H. Loh.Implementation of a closed-loop structural control systemusing wireless sensor networks. Structural Control andHealth Monitoring, 15(4):518–539, 2011.

[23] Y. Ni, B. Li, K. Lam, D. Zhu, Y. Wang, J. Lynch, andK. Law. In-construction vibration monitoring of a super-tallstructure using a long-range wireless sensing system. SmartStructures and Systems, 7(2):83–102, March 2011.

[24] M. Pajic, S. Sundaram, J. Ny, G. Pappas, and R. Mangharam.Closing the loop: A simple distributed method for controlover wireless networks. In IPSN, 2012.

[25] J. Polastre, R. Szewczyk, and D. Culler. Telos: enablingultra-low power wireless research. In Proceedings of the 4thinternational symposium on Information processing insensor networks, IPSN ’05, Piscataway, NJ, USA, 2005.IEEE Press.

[26] A. Saifullah, P. Tiwari, B. Li, C. Lu, and Y. Chen.Accounting for failures in delay analysis for wirelesshartnetworks. In Tech. Rep., WUCSE-2012-16, WashingtonUniversity in St Louis, 2012.

[27] A. Saifullah, Y. Xu, C. Lu, and Y. Chen. Real-timescheduling for WirelessHART networks. In RTSS, 2010.

[28] B. Spencer, S. Dyke, and H. Deoskar. Benchmark problemsin structural control: part i-active mass driver system.Earthquake Engineering and Structural Dynamics,

27(11):1127–1139, 1998.[29] K. Srinivasan, P. Dutta, A. Tavakoli, and P. Levis. An

empirical study of low power wireless. ACM Transactions onSensor Networks, 2010.

[30] Z. Sun, B. Li, D. Dyke, and C. Lu. Evaluation ofperformances of structural control benchmark problem withtime delays from wireless sensor network. In Proc.EMI/PMC 2012, 2012.

[31] Z. Sun, B. Li, D. Dyke, and C. Lu. A novel data utilizationand control strategy for wireless structural control systemswith tdma network. In Proc. ASCE IWCCE 2013, 2013.

[32] Y. Wang and K.-H. Law. Structural control with multi-subnetwireless sensing feedback: experimental validation oftime-delayed decentralized hr control design. Advances inStructural Engineering,, 14(1):25–39, 2011.

[33] Y. Wang, R. Swartz, J. Lynch, K. Law, and C.-H. Loh.Performance evaluation of decentralized wireless sensingand control in civil structures. In Proceedings of SPIE 14thInternational Symposium on Smart Structures and Materialsand Nondestructive Evaluation and Health Monitoring,March 2007.