Decentralized Civil Structural Control using Real-time Wireless Sensing and Embedded Computing

Transcription

1 Decentralize Civil Structural Control using Real-time Wireless Sensing an Embee Computing Yang Wang a, R. Anrew Swartz b, Jerome P. Lynch * b, Kincho H. Law a, Kung-Chun Lu c, Chin-Hsiung Loh c a Dept. of Civil an Environmental Engineering, Stanfor Univ., Stanfor, CA 94305, USA b Dept. of Civil an Environmental Engineering, Univ. of Michigan, Ann Arbor, MI 4809, USA c Dept. of Civil Engineering, National Taiwan Univ., Taipei, Taiwan * Corresponence Author: Jerome P. Lynch Assistant Professor Department of Civil an Environmental Engineering The University of Michigan 2380 G. G. Brown Builing Ann Arbor, MI USA jerlynch@umich.eu Abstract Structural control technologies have attracte great interest from the earthquake engineering community over the last few ecaes as an effective metho of reucing unesire structural responses. Traitional structural control systems employ large quantities of cables to connect structural sensors, actuators, an controllers into one integrate system. To reuce the high-costs associate with labor-intensive installations, wireless communication can serve as an alternative real-time communication link between the noes of a control system. A prototype wireless structural sensing an control system has been physically implemente an its performance verifie in large-scale shake table tests. This paper introuces the esign of this prototype system an investigates the feasibility of employing ecentralize an partially ecentralize control strategies to mitigate the challenge of communication latencies associate with wireless sensor networks. Close-loop feeback control algorithms are embee within the wireless sensor prototypes allowing them to serve as controllers in the control system. To valiate the embement of control algorithms, a 3-story half-scale steel structure is employe with magnetorheological (MR) ampers installe on each floor. Both numerical simulation an experimental results show that ecentralize control solutions can be very effective in attaining the optimal performance of the wireless control system. Key wors: structural control, wireless communication, embee computing, ecentralize control, velocity feeback control

2 . Introuction As an effective metho of reucing the ynamic response of structures uring earthquakes or typhoons, structural control technologies have attracte a great amount of interest from structural engineering researchers an practitioners over the past few ecaes (Soong an Spencer, 2002). About 50 builings an towers were instrumente with various types of structural control systems from 989 to 2003 (Chu et al., 2005), with evient reuction in structural ynamic responses being reporte. Current structural control systems can be categorize into three major types: (a) passive control (e.g. base isolation), (b) active control (e.g. active mass ampers), an (c) semiactive control (e.g. semi-active variable ampers). Passive control has the avantage of power efficiency, while active control has the avantage of being aaptable to real-time excitations. As a hybri between these two approaches to structural control (active an passive), semi-active control effectively combines the avantages of both systems. Examples of semi-active actuators inclue active variable stiffness (AVS) evices, semi-active hyraulic ampers (SHD), electrorheological (ER) ampers, an magnetorheological (MR) ampers. Another attractive feature of semi-active control systems is that they are inherently stable because their actuators o not apply mechanical energy irectly to the structure. In both semi-active an active control systems, sensors are employe in the structure to collect real-time structural response ata uring ynamic excitation (e.g. win or earthquake). The response ata is then fe into a single or multiple control ecision moules (controllers) in orer to etermine control forces an apply commans to system actuators in real-time. Accoring to these comman signals, the structural actuators generate control forces intene to reuce unwante structural vibrations. In traitional semi-active or active control systems, coaxial wires are normally use to provie communication links between sensors, actuators an controllers. For a typical low-rise builing, the installation of a commercial wire-base ata acquisition (DAQ) system can cost upwars of a few thousan ollars per sensing channel (Celebi, 2002). As the 2

3 size of the control system grows (efine by the noal ensity an inter-noal spatial istances), the aitional cabling neee may result in increases in installation time an expense (Solomon et al., 2000). To capitalize on future low-cost semi-active evices that are ensely installe in a structure, wireless communication technology can be aopte to eraicate the coaxial cables associate with traitional control systems. Although wireless communication has been wiely explore for use in structural monitoring applications (Straser an Kiremijian, 998; Lynch an Loh, 2006a; Wang et al., 2007), application to real-time feeback control in structural engineering has been scarce (Lynch an Tilbury, 2005). Outsie of structural engineering, a few examples of wireless control systems have been reporte (Eker et al., 200; Ploplys et al., 2004). When replacing wire communication channels with wireless ones for feeback structural control, ifficulties inclue coorination of the wireless noes in a collaborative control network, egraation of real-time performance, an higher probability of ata loss uring transmission. The egraation of the control system s real-time characteristics is a common problem face by istribute network control systems, regarless of using wire or wireless communication (Lian et al., 2002). Among the ifferent solutions propose for this problem, one possible remey is the aoption of ecentralize control strategies (Sanell et al., 978; Lynch an Law, 2002). In a ecentralize control system, the sensing an control network is ivie into multiple subsystems. Controllers are assigne to each subsystem an require only subsystem sensor ata to make control ecisions. Therefore, reuce use of the communication channel is offere by a ecentralize control architecture, which results in higher control sampling rates. Furthermore, ecentralize control requires relatively shorter communication ranges, enabling more reliable wireless ata transmissions. The rawback of ecentralize control is that ecentralize system architectures may only achieve sub-optimal control performance compare with centralize counterparts, because each subsystem only has its own state ata from which it must calculate 3

4 control ecisions. This work attempts to investigate the effectiveness of ecentralize wireless control in civil structures. This paper first introuces a prototype wireless structural sensing an control system evelope by the authors (Wang et al., 2006). The system consists of multiple stan-alone wireless sensors that form an integrate network through a common wireless communication channel. Each wireless sensor can recor response ata from sensors, calculate control forces, communicate state ata, an comman actuators. In orer to investigate the effects of communication latencies on centralize an ecentralize wireless control strategies, centralize an ecentralize output feeback control algorithms are implemente. The ecentralize control metho is evelope by applying appropriate shape constraints to the gain matrix of a normal optimal output feeback control problem. Numerical simulations show that the higher sampling rates achievable by ecentralize control may compensate for the isavantage of incomplete sensor ata from which control ecisions are mae. Large-scale shake table experiments are conucte on a 3-story steel frame structure installe with MR ampers to compare the performance of ifferent ecentralize an centralize control schemes. 2. A Prototype Real-time Wireless Sensing an Control System To illustrate the architecture of the prototype wireless sensing an control system, Fig. shows a 3-story structure controlle by three actuators. Wireless sensors an controllers are mounte on the structure for measuring structural response ata an commaning actuators in real-time. Besies the wireless sensing an control units that are necessary for ata collection an the operation of the actuators, a remote comman server with a wireless transceiver is inclue in the system to initiate the operation of the control system an to log the flow of wireless ata. To initiate the operation of the control system, the comman server first broacasts a start signal to all of the wireless sensing an control units. Once the start comman is receive, the wireless 4

5 units that are responsible for collecting sensor ata start acquiring an broacasting ata at a preset time interval. Accoringly, the wireless units responsible for commaning the actuators receive the sensor ata, calculate esire control forces, an apply control commans within the specifie time interval allotte at each time step. The following escribes in etail the esign of the wireless sensing an control units. Specific attention is pai to the control signal generation an wireless communication moules of the wireless unit since both are integral to the performance of the global control system. 2. Harware Design for the Wireless Sensing an Control Unit The wireless unit is esigne in such a way that the unit can serve as either a sensing unit (i.e. a unit that collects ata from sensors an wirelessly transmits the ata), a control unit (i.e. a unit that calculates control forces an commans actuators), or a unit for both sensing an control. This flexibility is supporte by an integrate harware esign base upon a wireless sensing unit previously propose for wireless structural monitoring by Wang et al. (2007). Three basic functional moules are inclue in the wireless sensing unit esign: sensor signal igitizer, computational core, an wireless transceiver. The wireless control unit contains the same three moules as the wireless sensing unit, but also inclues a supplementary control signal generation moule esigne to resie off-boar of the basic wireless sensor. The architectural esign of the wireless control unit is presente in Fig. 2(a). The architectural esign of the wireless sensing unit can be obtaine by simply omitting the control signal generation moule. A simple twolayer printe circuit boar (PCB) for the wireless sensing unit is esigne an fabricate (Fig. 2b). As shown in Fig. 2(c), the PCB, wireless transceiver, an batteries are store within an off-theshelf weatherproof plastic container, which has the imensions of 0.2cm by 6.5cm by 4.0cm. The sensor signal igitization moule of the wireless unit contains a 4-channel 6-bit Texas Instrument ADS834 analog-to-igital (A/D) converter. This moule converts the 0 to 5V analog 5

6 output of up to four sensors (to ate, ifferent types of transucers have been successfully interface incluing accelerometers, velocity meters, LVDTs an strain gages) into igital formats usable by the wireless sensor s computational core. The igitize sensor ata is transferre to the computational core through a high-spee serial peripheral interface (SPI) port. The computational core consists of a low-power 8-bit Atmel ATmega28 microcontroller an an external 28kB static ranom access memory (SRAM) chip for ata storage. For wireless communication among the sensing units, two types of wireless transceivers may be employe: 900MHz MaxStream 9XCite an 2.4GHz MaxStream 24XStream (MaxStream, 2005). Generally, the 9XCite is use in the U.S. an other countries where 900MHz is open to unlicense usage while the 24XStream is use internationally on the 2.4GHz raio ban. When a wireless unit is esignate as a sensing unit, its wireless transceiver is primarily use to sen sensor ata out to the wireless network. In contrast, for a wireless control unit, its wireless transceiver receives sensor ata from the network. After receiving the sensor ata, the ATmega28 microcontroller embee in the wireless control unit computes the esire control forces. Once the control force calculation is complete, the wireless control unit issues commans (voltage signals) to the actuator through the control signal generation moule. All of the harware components of the wireless sensing unit, not incluing the wireless transceiver, consume about 32mA when active an 80µA when in stan-by moe. The aitional power consumption of the two wireless transceivers (9XCite an 24XStream) will be presente in section Control Signal Generation Moule A separate harware moule is esigne to be connecte with the wireless sensing unit, which permits the unit to generate analog voltage signals for commaning actuators. At the core of this control signal generation moule is the single-channel 6-bit igital-to-analog (D/A) converter, the Analog Device AD5542. The AD5542 receives a 6-bit unsigne integer from the ATmega28 an converts the integer value to an analog voltage output spanning from -5 to 5V. 6

7 Aitional support electronics are inclue in the control signal generation moule to offer stable zero-orer hol voltage outputs at high sampling rates ( MHz maximum). The wie voltage output range (-5 to 5V) of the control signal generation moule, particularly the negative output range, is one of the key features of the moule s esign. With the wireless sensing unit esigne to operate on 5V, the Texas Instruments PT5022 switching regulator is integrate in the signal generation moule to convert the 5V regulate power supply into a stable -5V reference. Another auxiliary component require for the AD5542 to generate a bipolar -5 to 5V output signal is a rail-to-rail input an output operational amplifier, for which the National Semiconuctor LMC6484 operational amplifier is selecte. The typical slew rate of the LMC6484 is about.3v/µs, which means that the output voltage can swing about.3v within µs. This output graient is compatible with the microsecon-level settling time of the AD5542 D/A converter. When operational, the control signal generation moule raws about 70mA from the 5V power supply provie by the wireless sensing unit. As shown in Fig. 3(a), a separate ouble-layer printe circuit boar (PCB) is esigne to accommoate the D/A converter (AD5542) an its auxiliary electrical components. The control signal boar is attache via two multi-line wires to the wireless sensor. To reuce circuit noise, two separate wires are use with one eicate to analog signals an the other to igital signals. The analog signal cable transfers an accurate +5V reference voltage from the existing wireless sensing boar to the signal generation moule. The igital signal cable provies all of the connections require to accommoate the serial peripheral interface (SPI) between the ATmega28 microcontroller an the AD5542. To comman an actuator, a thir wire is neee to connect the output of the control signal generation moule with the structural actuator. The control signal generation moule connecte to the wireless sensing unit is shown in Fig. 3(b). 2.3 Wireless Communication Moule 7

8 A challenge associate with employing wireless sensors in a structural control system is the performance of the wireless communication channel. When the wireless sensors are use within a wireless structural monitoring system, robust sen-acknowlegement communication protocols are use to ensure that no ata is lost (Lynch et al., 2006b). However, the real-time requirements of the control system o not permit sufficient time for the use of sen-acknowlegement communication protocols that ensure channel reliability. Furthermore, stochastic elays in the channel are probable; these elays cannot be eterministically accounte for a priori in the control solution formulation (Seth et al., 2004). For these reasons, an appropriate wireless transceiver that minimizes ata loss without sacrificing communication spee must be juiciously selecte for use in the wireless control system. The wireless sensing unit is esigne to be operable with two ifferent wireless transceivers: 900MHz MaxStream 9XCite an 2.4GHz MaxStream 24XStream. Pin-to-pin compatibility between these two wireless transceivers makes it possible for them to share the same harware connections in the wireless unit. Because of the ifferent ata rates, embee software for using the two transceivers is slightly ifferent. This ual-transceiver support offers the wireless sensing an control unit more flexibility in terms of not only use in ifferent geographical areas, but also provies ifferent ata transfer rates, communication ranges, an power consumption characteristics. Table summarizes the key performance parameters of the two wireless transceivers. As shown in the table, the ata transfer rate of the 9XCite is twice that of the 24XStream, while the 24XStream provies a longer communication range but consumes more battery power. The peer-to-peer communication capability of the two wireless transceivers makes it possible for the wireless sensing an control units to communicate with each other, thus supporting flexible information flow among multiple wireless units. In this stuy, valiation tests of the wireless sensing an control system are performe using a test structure at the National Center for Research on Earthquake Engineering (NCREE) in Taipei, Taiwan. Because of the 8

9 local frequency ban requirements in Taiwan, the MaxStream 24XStream wireless transceiver operating on the 2.4GHz spectrum is employe for the experimental tests. As previously iscusse, one critical issue in applying wireless communication technology to real-time feeback structural control problems is the communication latency encountere when transmitting sensor ata from the wireless sensing units to the wireless control units. The anticipate transmission time of the 24XStream raio for a single ata packet is illustrate in Fig. 4. The transmission time consists of the communication latency (T Latency ) of the raios an the time to transfer ata between the microcontroller an the raio using the universal asynchronous receiver an transmitter (UART) interface (T UART ). Assume that the ata packet to be transmitte contains N bytes an the UART ata rate is R UART bps (bits per secon), which is equivalent to R UART /0 bytes per secon, or R UART /0000 bytes per millisecon. It shoul be note that the UART transmits 0 bits for every one byte (8-bits) of sensor ata ue to start an stop bits. The communication latency in a single transmission of this ata packet can be estimate as: T SingleTransm 0000N = TLatency + (ms) () R UART In the prototype wireless sensing an control system, the setup parameters of the 24XStream transceiver are first tune to minimize the transmission latency, T Latency. Then experiments are conucte to measure the actual T Latency, which turns out to be 5±0.5ms. The UART ata rate of the 24XStream raio, R UART, can be selecte among seven ifferent options, ranging from 200 bps to bps. After consiering the UART performance of the ATmega28 microcontroller operate at 8MHz, R UART is selecte as bps in the implementation. If a ata packet sent from a sensing unit to a control unit contains bytes, the total time elay for a single transmission is estimate to be: 9

10 T SingleTransm 0000 = (ms) (2) This single-transmission elay represents the communication constraint that nees to be consiere when calculating the upper boun for the maximum sampling rate for the control system. Another eciing factor for the sampling rate is the number of wireless transmissions neee at each control sampling step. A few millisecons of safety cushion time at each sampling step is a pruent aition that allows a certain amount of ranomness in the wireless transmission latency without unermining the reliability of the communication system. It shoul be note that the transmission latency, T Latency, for a MaxStream 9XCite transceiver can be as low as 5ms. This lower latency makes the 9XCite transceiver more suitable for real-time feeback control applications compare with the 24XStream transceiver. However, the 9XCite transceiver may only be use in countries where the 900MHz ban is for free public usage, such as the U.S., Canaa, Mexico, South Korea, an Japan. 3. Centralize an Decentralize Time-elay Control Algorithms using Velocity Feeback An optimal feeback control esign normally requires aequate real-time structural response ata to compute optimal control forces. For example, if a multi-story builing is moele by a lumpe-mass structural system with actuators eploye among ajacent floors, real-time floor isplacements an velocities that constitute the state-space vector are neee for a typical linear quaratic regulator (LQR) controller (Franklin et al., 2003). However, ue to instrumentation complexity an cost, not all structural response ata may be available in practice. To aress this ifficulty, output feeback control methos can be use to provie an optimal control strategy uner the constraint that only part of the state-space variables are measure in real-time. This section first presents the basic formulation of an optimal centralize output feeback control 0

11 solution an then proposes a moifie algorithm that allows the output feeback gain matrix to be constraine. The output feeback gain matrix is then formulate for various ecentralize control architectures using the constraine gain matrix algorithm etaile herein. 3. Formulation for Centralize Linear Output Feeback Control The output feeback igital-omain LQR control solution can be briefly summarize as follows. For a lumpe-mass structural moel with n egrees-of-freeom (DOF) an m actuators, the system state-space equation consiering l time steps of elay can be state as: z [ k+ ] = A z [ k] + B p [ k l], where z [ k] [ k] [ k] x = x (3) Here, z [ k] represents the 2n iscrete-time state-space vector, [ k] base) isplacement of the structural egrees-of-freeom, [ k l] force vector, x is the relative (to the p is the elaye m control A is the 2n 2n system matrix (containing the information about structural mass, stiffness an amping), an B is the 2n m actuator location matrix. The primary objective of the time-elay LQR problem is to minimize a global cost function, J, by selecting an optimal control force trajectory p : T T ( [ ] [ ] [ ] [ ]) 2 2 J = z k Qz k + p k l Rp k l, where Q 0 an R > 0 (4) p k= l n n m m In an output feeback control esign, when control ecisions are compute, only ata in the system output vector [ k] y are available. The output vector is efine by a q 2n linear transformation, D, to the state-space vector [ k] z :

12 [ k] = [ k] y D z (5) For example, if the relative velocities on all floors are measurable but no relative isplacement is measurable, D can be efine as: [ ] D = 0 I (6) _cen n n n n In another example, if only inter-story velocities between ajacent floors are measurable, the following output matrix D can be use: D _ec = (7) The m q optimal gain matrix G is require to provie a linear output feeback control: [ k] = [ k] p G y (8) Chung et al. (995) proposes a solution to the output feeback control problem with time elay (say, l time steps) by introucing a moifie first-orer ifference equation: [ k + ] = [ k] + [ k] z A z B p (9) in which the augmente state [ k] z is assemble from the past l states as: 2

13 z [ k] [ k] [ k ] z = z z [ k l] nl ( ) 2 + (0) This system is equivalent to the original system (Eq. 3) by proper efinitions of the augmente matrices an vectors: A an B are the augmente state-space system matrices, D is the augmente output matrix, Q is the augmente weighting matrix, an Z l is the secon statistical moment of the augmente initial isturbance. As a result, the following nonlinearly couple matrix equations are simultaneously solve for an optimal output feeback gain matrix Lagrangian matrix, L, an a constant matrix, H: G, the T T T ( ) ( ) ( ) A + BGD HA + BGD H+ Q+ D G RGD = 0 (a) ( ) ( ) T l A + B G D L A + B G D L+ Z = 0 (b) ( ) T T T 2B H A + B G D LD + 2RG D LD = 0 (c) Intereste reaers are referre to Chung et al. (995), where the time-elay optimal control solution is erive in etail. 3.2 Heuristic Solution for Centralize an Decentralize Output Feeback Gain Matrices An iterative algorithm to solve the continuous-time feeback control problem has been presente by Lunze (990). The algorithm (Fig. 5) starts from an initial estimate for the gain matrix Within each iteration step i, the matrix H i an L i are solve respectively using the current estimate of the gain matrix G. G i. Base on the H i an L i matrices compute, a searching graient 3

14 is calculate an the new gain matrix G is compute by traversing along a graient from i i+ G i. An aaptive multiplier, s, is use to ynamically control the search step size. At each iteration step, two conitions are use to ecie whether G i+ is an acceptable estimate. The first conition is trace( Hi Z l) < trace( HZ i l) which guarantees that G i+ is a better solution than + G i. The secon conition is that the maximum magnitue of all the eigenvalues of the matrix ( + i+ ) A B G D is less than which ensures the stability of the augmente system. The iterative algorithm put forth by Lunze (990) has an inherently powerful an attractive feature; the algorithm can be use to formulate an optimal control solution for a ecentralize system simply by constraining the structure of G to be consistent with the ecentralize architecture. The following equation presents the structure of two ecentralize output feeback gain matrices for a simple 3-story lumpe-mass structure. * 0 0 * * 0 G 0 * 0, an 2 0 * * _ec = G _ec = 0 0 * 0 * * (2) The pattern in G _ec specifies that when computing control ecisions, the actuator on each floor only nees the entry in the output vector y that correspons to that floor. The pattern in G _ec2 specifies that information from neighboring floors be consiere in calculating control actions. In orer to fin a ecentralize gain matrix that satisfies esire architectural constraints, the algorithm escribe in Fig. 5 is moifie by zeroing out the entries in the graient matrix i corresponing to zero terms in the ecentralize output feeback gain matrix. The next estimate G i+ is compute by traversing along this constraine graient. Using the above ecentralize 4

15 gain matrices an the output matrix D _ec efine in Eq. (7), inter-story velocities between ajacent floors can be use for the calculation of control actions in a ecentralize control system. 3.3 Simulation Results using Centralize an Decentralize Control Strategies Numerical simulations have been conucte to assess the performance of ecentralize an centralize control strategies consiering time elays ue to communication latencies. A numerical moel for the 3-story half-scale laboratory structure illustrate in Fig. is use for the simulation. For simplicity, ieal structural actuators which are capable of proucing any esire force uner the maximum limit of 20kN is eploye between each pair of ajacent floors using the V-braces shown. Three control architectures are employe: () ecentralize, (2) partially ecentralize, an (3) centralize. Different patterns of the gain matrices, matrices, G, an the output D, for these three control architectures are summarize in Table 2. As efine by these matrices, one centralize an two ecentralize velocity feeback patterns are aopte. An LQR weighting matrix Q minimizing inter-story rifts over time an a iagonal weighting matrix R are use when esigning the optimal gain matrices for all the simulations presente herein. Various combinations of centralization egrees (: fully ecentralize; 2: partially ecentralize; 3: centralize) an sampling time steps ranging from 0.005s to 0.s (at a resolution of 0.005s) are simulate. To assess the performance of each control scheme, three groun motion recors are use for the simulation: 940 El Centro NS (Imperial Valley Irrigation District Station), 999 Chi-Chi NS (TCU-076 Station), an 995 Kobe NS (JMA Station) earthquake recors. Performance inices propose by Spencer et al. (998) are aopte. In particular, two representative performance inices employe are: 5

16 PI max () t () t i ti, = max El Centro ˆ Kobe max i Chichi ti, J LQR, an PI2 = max Jˆ LQR El Centro Kobe Chichi (3) where PI an PI 2 are the performance inices corresponing to inter-story rifts an LQR cost inices, respectively. In Eq. (3), i () t represents the inter-story rift between floor i (i =, 2, 3) an its lower floor at time t, an max i () t is the maximum inter-story rift over the entire time ti, history an among all three floors. The maximum inter-story rift is normalize by its counterpart max ˆ i () t, the maximum response of the uncontrolle structure. The largest ti, normalize ratio among the simulations for the three ifferent earthquake recors is efine as the performance inex PI. Similarly, the performance inex PI 2 is efine for the LQR control inex J LQR, as given in Eq. (4). When computing the LQR inex over time, a uniform time step of 0.005s is use to collect the structural response ata points, regarless of the sampling time step of the control scheme; this allows one control strategy to be compare to another without concern for the ifferent sampling time steps use in the control solution. Values of the two control performance inices are plotte in Fig. 6 for ifferent combinations of centralization egrees an sampling time steps. The plots shown in Fig. 6(a) an 6(b) illustrate that centralization egree an sampling step have significant impact on the performance of the control system. Generally speaking, control performance is better for higher egrees of centralization an shorter sampling times. To better review the simulation results, the performance inices for the three ifferent control schemes are re-plotte as a function of sampling time in Fig. 6(c) an 6(). As shown in Fig. 6(c), if a partially ecentralize control system can achieve 0.04s sampling step an a centralize system can only achieve 0.08s ue to aitional communication latency, the partially ecentralize system can result in lower 6

17 maximum inter-story rifts. Similar trens are observe in Fig. 6(), although for a given sampling time step, the performance inex PI 2 for the centralize case is always lower than the inices for the two ecentralize cases. 4. Valiation Experiments using a 3-story Structure Instrumente with MR Dampers To stuy the potential use of the wireless sensing an control system for ecentralize structural control, valiation tests are conucte at the National Center for Research on Earthquake Engineering (NCREE) in Taipei, Taiwan. The same test structure has been use in a previous wireless control stuy in which one MR amper is installe for close-loop control (Lynch et al. 2006c). Both a baseline wire control system an a wireless sensing an control system are employe to implement the real-time feeback control of a 3-story steel frame instrumente with three MR ampers. 4. Valiation Test Setup A three-story steel frame structure is esigne an constructe by researchers affiliate with NCREE (Fig. 7a). The imensions of the structure are provie in Fig.. The three-story structure is mounte on a 5m 5m 6-DOF shake table. The shake table can generate groun excitations with frequencies spanning from 0.Hz to 50Hz. For this stuy, only longituinal excitations are use. Along this irection, the shake table can excite the structure with a maximum acceleration of 9.8m/s 2. The excitation has a maximum stroke an force of ±0.25m an 220kN, respectively. The test structure an shake table are heavily instrumente with accelerometers, velocity meters, an linear variable isplacement transucers (LVDT) to measure their ynamic response. These sensors are interface to a high-precision wire-base ata acquisition (DAQ) system permanently installe in the NCREE facility; the DAQ system is set to a sampling rate of 200 Hz. A separate set of wireless sensors are installe as part of the wireless control system. 7

18 For the wireless system, a total of four wireless sensors are installe following the eployment strategy shown in Fig.. Each wireless sensor is interface to a Tokyo Sokushin VSE5-D velocity meter to measure the absolute velocity response of each floor of the structure as well as at the base (i.e. table top velocity). The sensitivity of the velocity meter is 0V/(m/s) with a measurement limit of ± m/s. The three wireless sensors on the first three levels of the structure (C 0, C, an C 2 ) are also responsible for commaning the MR ampers. Besies the wireless control system, a traitional wire-base control system is installe in the structure for comparative analyses. Centralize an ecentralize velocity feeback control schemes escribe earlier (Table 2) are use for both the wire an the wireless control systems. As shown in Table 3, ifferent ecentralization patterns an sampling steps are teste. For the test structure, the wire-base system can achieve a sampling rate of 200Hz, or a time step of 0.005s. Mostly ecie by the communication latency of the 24XStream wireless transceivers, the wireless system can achieve a sampling rate of 2.5Hz (or a time step of 0.08s) for the centralize control scheme. This sampling rate is ue to each wireless sensor waiting in turn to broacast its ata to the network (about 0.02s for each transmission). An avantage of the ecentralize architecture is that fewer communication steps are neee, thereby reucing the time for wireless communication. As shown in Table 3, the wireless system can achieve a sampling rate of 6.67Hz for partially ecentralize control an 50Hz for fully ecentralize control. 4.2 Magnetorheological (MR) Damper Hysteresis Moel For this experimental stuy, three 20 kn MR ampers are installe with V-braces on each story of the steel structure (Fig. 7b). The amping coefficients of the MR ampers can be change in real-time by the wireless control units (Fig. 7(c)) simply by issuing a comman voltage between 0 to.2v. A separate power moule is neee to provie 24V of power to each MR amper. In aition, this power moule takes the comman voltage as an input an converts this comman 8

19 signal to a regulate current from 0 to 2A. The current is input to the internal electromagnetic coil of the MR amper to generate a magnetic fiel that sets the viscous amping properties of the MR flui containe within the amper cyliner. Therefore, the axial force require to move the amper piston against the cyliner is ajustable through the comman voltage. Calibration tests are first conucte on the MR ampers before mounting them to the structure so that moifie Bouc-Wen amper moels can be formulate for each amper (Lin et al., 2005). The nonlinear force-velocity relationship for this moifie Bouc-Wen moel is efine as: () = ( ) () + () F t C V x t z t (4) where F(t) is the force provie by the MR amper, C(V) is the amping coefficient ajustable through the amper comman voltage V, x () t is the relative velocity of the amper piston against the cyliner, an z(t) is the hysteresis restoring force of the amper. The hysteresis restoring force, z(t), evolves accoring to the following ifferential equation (Lin et al., 2005): N n () = () + n β () () () + γ () () n z t Ax t a x t z t z t x t z t (5) n= where A, β, γ, a n, an N are parametric constants of the Bouc-Wen moel. The higher the orer N, the more accurate the moel is in escribing the complicate hysteresis behavior of the semiactive amper. In this stuy, it was foun that an orer of N = 2 provies fairly accurate moeling to the amper hysteresis forces. The hysteresis restoring force in iscrete-time form at step k is then rewritten as: 9

20 [ ] = [ ] + T [ ] ( ) zk zk Φ k Θ V (6a) ( V) = { θ( V) θ2( V) θ3( V) θ4( V) θ5( V) } T Θ (6b) { } [ k] = t xk [ ] xk [ ] zk [ ] xk [ ] zk [ ] xk [ ] zk [ ] zk [ ] xk [ ] zk [ ] Φ 2 T (6c) The five parameters in vector Θ ( V ) are moele as low-orer polynomial functions of the amper voltage V. Similarly, the amping coefficient, C(V), is moele as a linear function of the amper voltage V. For each MR amper, the constant coefficients in the polynomial functions Θ ( V ) an C(V) are pre-etermine an valiate through a series of calibration experiments with the amper (Lin et al., 2005). In these experiments, a isplacement-controlle actuator is employe to comman isplacements to the amper piston while the amper cyliner is fixe to a reaction frame. A loa cell is then use to measure the amper force, so that the forceisplacement time histories of the amper can be recore. In orer to compute the constant coefficients in the moifie Bouc-Wen moel, sinusoial an ranom isplacement are first applie to the amper with the amper comman voltage fixe at multiple levels. Then the moel is valiate through experiments when ranom isplacement time histories are applie to the amper with the comman voltage ranomly varie. In the real-time feeback control tests, hysteresis status upating for the MR ampers is an integral element in the calculation of amper actuation voltages. At each sampling time step, a wireless control unit first ecies the esire control force for the MR amper using the control algorithms escribe in Section 3 (Eq. 8). Meanwhile, the unit calculates the amper hysteresis status accoring to the moifie Bouc-Wen moel (Eq. 6). Accoring to the hysteresis status, the wireless control unit ecies the appropriate comman voltage neee to be applie to the MR amper to attain a amping force closest to the esire control force. Fig. 8 illustrates the comparison between the control force esire by the wireless control units an the actual force 20

21 (as measure by the loa cells) achieve by the MR ampers on Floor-0 an Floor- uring a centralize control test. The groun excitation in this test is the 940 El Centro NS earthquake recor scale to a peak groun acceleration of m/s 2. The strong similarity between the esire an achieve control forces valiates the amper hysteresis computation accomplishe by the wireless control units, an the effectiveness of the moifie Bouc-Wen MR amper moel. 4.3 Experimental Results To ensure that appropriate control ecisions are compute by the wireless control units, one necessary conition is that the real-time velocity ata use by the control units are reliable. Rarely experiencing ata losses uring the experiments, our prototype wireless sensor network proves to be robust (as reporte by Lynch et al. (2006c), ata losses less than 2% are experience). Shoul ata loss be encountere, the wireless control unit is currently esigne to simply use a previous ata sample. To illustrate the reliability of the velocity ata collecte an transmitte by the wireless units, Fig. 9(a) presents the Floor- time history ata uring the same centralize wireless control test as presente in Fig. 8. The ata is collecte separately by the cable DAQ system an recore by the three wireless control units. During the test, unit C measures the ata from the associate velocity meter irectly, stores the ata in its own memory bank, an transfers the ata wirelessly to unit C 0 an C 2. After the test run is complete, ata from all the three control units are sequentially streame to the experiment comman server, where the results are plotte as shown in Fig. 9(a). These plots illustrate strong agreements among ata recore by the three wireless control units an by the cable system using a separate set of velocity meters an ata acquisition system. This result shows that the velocity ata is not only reliably measure by unit C 0, but also properly transmitte to the other wireless control units in real-time. The time histories of the inter-story rifts from the same centralize wireless control test are plotte in Fig. 9(b), together with the rifts of a centralize wire control test an a ynamic test 2

22 when the structure is not instrumente with any control system (i.e. the MR ampers are not yet installe). The same groun excitations (e.g. the 940 El Centro NS earthquake recor scale to a peak groun acceleration of m/s 2 ) are use for all the three cases shown in Fig. 9(b). The results show that both the wireless an wire control systems achieve consierable gain in limiting inter-story rifts. Running at a much shorter sampling time step, the wire centralize control system achieves slightly better control performance than the wireless centralize system in terms of mitigating inter-story rifts. To further stuy ifferent ecentralize schemes with ifferent communication latencies, Fig. 0 shows the peak inter-story rifts an floor accelerations for the original uncontrolle structure an the structure controlle by the four ifferent wireless an wire control schemes, as efine in Table 3. Three earthquake recors, the 940 El Centro NS, 999 Chi-Chi NS, an 995 Kobe NS recors, are employe for the experimental tests, with their peak groun accelerations all scale to m/s 2. Compare with the uncontrolle structure, all wireless an wire control schemes achieve significant reuction with respect to maximum inter-story rifts an absolute accelerations. Among the four control cases, the wire centralize control scheme shows goo performance in mitigating both peak rifts an peak accelerations. For example, when the 940 El Centro NS earthquake is employe (Fig. 0a), the wire centralize control scheme achieves the smallest peak rifts an secon smallest overall peak accelerations. This result is expecte because the wire system has the avantages of lower communication latency an utilizes sensor ata from all floors (complete state ata). The wireless schemes, although running at longer sampling steps, achieve control performance comparable to the wire system. For all three earthquake recors, the fully ecentralize wireless control scheme (case #) results in low peak inter-story rifts an the smallest peak floor accelerations at most of the floors. This result illustrates that in the ecentralize wireless control cases, the higher sampling rate (achieve ue 22

23 to lower communication latency) potentially compensates for the lack of ata available since sensor ata from faraway floors is ignore. 5. Summary an Conclusions This paper investigates the feasibility an effectiveness of ecentralize wireless control strategies in civil structures. The aoption of wireless telemetry for structural control applications is avantageous because it reuces the nee for wiring between sensors, actuators an controllers yet it offers flexible communication architectures with moifiable network topologies. A prototype wireless structural sensing an control system esigne for real-time civil structural control is first introuce. We then present the theoretical backgroun for an optimal output feeback structural control esign using centralize an ecentralize communication patterns. Both numerical simulations an experimental tests are performe to examine the traeoff between the egree of centralization an communication latencies. The simulate an experimental results show that ecentralize wireless control strategies may provie equivalent or even superior control performance, given that their centralize counterparts suffer longer sampling steps ue to wireless communication latencies. Laboratory experiments also successfully valiate the reliability of the prototype wireless structural sensing an control system. With larger-scale control systems (efine by higher noal ensities) encountering greater communication complexities (i.e. communication elays, ata loss, an limite communication range), more work is neee to explore the traeoffs between egree of ecentralization, sample rate an global control system performance. Acknowlegments This research is partially fune by the National Science Founation uner grants CMS (Stanfor University), CMS (University of Michigan), an the Office of Naval Research Young Investigator Program aware to Prof. Lynch at the University of Michigan. Aitional 23

24 support is provie by National Science Council in Taiwan uner Grant No. NSC Z The authors wish to thank the two fellowship programs: the Office of Technology Licensing Stanfor Grauate Fellowship an the Rackham Grant an Fellowship Program at the University of Michigan. Reference Celebi, M. (2002), Seismic Instrumentation of Builings (with Emphasis on Feeral Builings), Report No , Unite States Geological Survey (USGS), Menlo Park, CA, USA. Chu, S.Y., Soong, T.T. an Reinhorn, A.M. (2005), Active, Hybri an Semi-active Structural Control, John Wiley & Sons Lt, West Sussex, Englan. Chung, L.L., Lin, C.C. an Lu, K.H. (995), Time-elay Control of Structures, Earthquake Engineering & Structural Dynamics, 24(5), Eker, J., Cervin, A. an Hörjel, A. (200), Distribute Wireless Control Using Bluetooth, Proc. of IFAC Conf. on New Technologies for Control System, Hong-Kong, China, November 9-22, 200. Franklin, G.F., Powell, J.D. an Workman, M. (2003), Digital Control of Dynamic Systems, Pearson Eucation, New Jersey. Lian, F.-L., Moyne, J. an Tilbury, D. (2002), Network Design Consieration for Distribute Control Systems, IEEE Transactions on Control Systems Technology, 0(2), Lin, P.-Y., Roschke, P.N. an Loh, C.-H. (2005). System Ientification an Real Application of a Smart Magneto-Rheological Damper, Proc. of the 2005 International Symposium on Intelligent Control, Limassol, Cyprus, June 27-29, Lunze, J. (992), Feeback Control of Large-scale Systems, Prentice Hall, Hertforshire, UK. Lynch, J.P. an Law, K.H. (2002). Decentralize Control Techniques for Large-scale Civil Structural Systems, Proc. of the 20th International Moal Analysis Conf., Los Angeles, CA, USA, February 4-7, Lynch, J.P. an Tilbury, D. (2005), Implementation of a Decentralize Control Algorithm Embee within a Wireless Active Sensor, Proc. of the 2n Annual ANCRiSST Workshop, Gyeongju, Korea, July 2-24, Lynch, J.P. an Loh, K. (2006a), A Summary Review of Wireless Sensors an Sensor Networks for Structural Health Monitoring, Shock an Vibration Digest, 38(2), Lynch, J.P., Wang, Y., Loh, K., Yi, J., an Yun, C.-B. (2006b), Performance Monitoring of the Geumang Brige using a Dense Network of High-Resolution Wireless Sensors, Smart Materials an Structures, IOP, 5(6), Lynch, J.P., Wang, Y., Swartz, R. A., Lu, K. C., an Loh, C. H. (2006c). Implementation of a Close-Loop Structural Control System using Wireless Sensor Networks, Journal of Structural Control an Health Monitoring, Wiley, in review. MaxStream, Inc. (2005). XStream OEM RF Moule Prouct Manual, Linon, UT, USA. 24

25 Ploplys, N.J., Kawka, P.A. an Alleyne, A.G. (2004), Close-loop Control over Wireless Networks, IEEE Control Systems Magazine, 24(3), Sanell, N., Jr., Varaiya, P., Athans, M. an Safonov, M., Survey of Decentralize Control Methos for Large Scale Systems, IEEE Transactions on Automatic Control, 23(2), Seth, S., Lynch, J.P. an Tilbury, D., Feasibility of Real-Time Distribute Structural Control upon a Wireless Sensor Network, Proceeings of the 42n Annual Allerton Conference on Communication, Control an Computing, Allerton, IL, USA, September 29 - October, Solomon, I., Cunnane, J. an Stevenson, P. (2000). Large-scale Structural Monitoring Systems, Proc. of SPIE Non-estructive Evaluation of Highways, Utilities, an Pipelines IV, Newport Beach, CA, March 7-9, Soong, T.T. an Spencer, B.F., Jr. (2002). Supplemental Energy Dissipation: State-of-the-art an State-of-the-practice, Engineering Structures, 24(3), Spencer, B.F., Jr., Christenson, R.E. an Dyke, S.J., (998). "Next Generation Benchmark Control Problem for Seismically Excite Builings." Proc. of 2n Worl Conf. on Structural Control, Kyoto, Japan, June 29 -July 2, 998. Straser, E.G. an Kiremijian, A.S. (998), A Moular, Wireless Damage Monitoring System for Structures, Report No. 28, John A. Blume Earthquake Eng. Ctr., Stanfor University, Stanfor, CA, USA. Wang, Y., Swartz, A., Lynch, J.P., Law, K.H., Lu, K.-C. an Loh, C.-H. (2006), Wireless Feeback Structural Control with Embee Computing, Proc. of the SPIE th International Symposium on Nonestructive Evaluation for Health Monitoring an Diagnostics, San Diego, CA, USA, February 26 - March 2, Wang, Y., Lynch, J.P. an Law, K.H. (2007), A Wireless Structural Health Monitoring System with Multithreae Sensing Devices: Design an Valiation, Structure an Infrastructure Engineering - Maintenance, Management an Life-Cycle Design & Performance, 3(2),

26 Key wors: structural control, wireless communication, embee computing, ecentralize control, velocity feeback control List of Figures an Tables Figure. Illustration of the prototype wireless sensing an control system using a 3-story structure controlle by three MR ampers. Figure 2. Wireless sensing unit: (a) Functional iagram etailing the harware esign of the wireless sensing unit interface with the actuation signal generation moule; (b) Printe circuit boar for the wireless sensing unit ( cm 2 ); (c) Package of the wireless sensing unit ( cm 3 ). Figure 3. Pictures of the control signal generation moule: (a) PCB boar ( cm 2 ); (b) Control signal generation moule connecte to wireless sensor. Figure 4. Communication latency of a single wireless transmission. Figure. 5. Heuristic algorithm solving the couple nonlinear matrix equations (Eq. 9) for centralize optimal time-elay output feeback control (Lunze, 990). Figure 6. Simulation results illustrating control performance inexes for ifferent sampling time steps an centralization egrees: (a) 3D plot for performance inex PI ; (b) 3D plot for performance inex PI 2 ; (c) Conense 2D plot for PI ; () Conense 2D plot for PI 2. Figure 7. Laboratory setup: (a) the 3-story test structure mounte on the shake table; (b) the MR amper installe between the st floor an the base floor of the structure; (c) a wireless control unit an an off-boar control signal generation moule. Figure 8. Damper forces esire by the control units an achieve by the MR ampers uring an experiment run: (a) force provie by the MR amper uner the st floor; (b) force provie by the MR amper uner the 2 n floor. Figure 9. Experimental time histories for: (a) Floor- absolute velocity ata recore by the cable an wireless sensing systems; (b) inter-story rifts of the structure with an without control. Figure 0. Experimental results of ifferent control schemes uner three earthquake excitations scale to peak groun accelerations of m/s 2 : (a) 940 El Centro NS; (b) 999 Chi-Chi NS; (c) 995 Kobe NS. Table. Key performance parameters of the wireless transceivers. Table 2. Different ecentralization patterns for the control simulations an experiments. Table 3. Different ecentralization patterns an sampling steps for the wireless an wire-base control experiments (egrees of centralization are efine as shown in Table 2). 26

27 S 3 V 3 Floor-3 C i : Wireless control unit (with one wireless transceiver inclue) 3m C 2 S i : Wireless sensing unit (with one wireless transceiver inclue) D 2 V 2 Floor-2 T i : Wireless transceiver D i : MR Damper V i : Velocity meter 3m D C V Floor- Floor plan: 3m x 2m Floor weight: 6,000kg Steel I-section beams an columns: H50 x 50 x 7 x 0 3m T Lab experiment 0 comman server D 0 C 0 V 0 Floor-0 Figure. Illustration of the prototype wireless sensing an control system using a 3-story structure controlle by three MR ampers. 27

28 Wireless Sensing Unit Computational Core Structural Sensors Sensor Signal Digitization 4-channel 6-bit Analog-to-Digital Converter ADS834 SPI Port 28kB External SRAM CY6228B Parallel Port 8-bit Microcontroller ATmega28 UART Port Wireless Communication Wireless Transceiver: 20kbps 2.4GHz 24XStream, or 40kbps 900MHz 9XCite SPI Port Control Signal Generation 6-bit Digital-to-Analog Converter AD5542 Structural Actuator (a) ATmega28 Microcontroller Connector to Wireless Transceiver Sensor Connector A2D Converter ADS834 Octal D-type Latch AHC573 SRAM CY6228B (b) (c) Figure 2. Wireless sensing unit: (a) functional iagram etailing the harware esign of the wireless sensing unit interface with the actuation signal generation moule; (b) printe circuit boar for the wireless sensing unit ( cm 2 ); (c) package of the wireless sensing unit ( cm 3 ). 28

29 Digital Connections to ATmega28 Microcontroller Integrate Switching Regulator PT5022 Analog Connections to ATmega28 Microcontroller Comman Signal Output Digital-to-Analog Converter AD5542 (a) Operational Amplifier LMC6484 (b) Figure 3. Pictures of the control signal moule: (a) PCB boar ( cm 2 ); (b) control signal moule connecte to wireless sensor. 29

30 Data packet sent from ATmega28 to 24XStream Sening Unit T Latancy T UART time Receiving Unit time Data packet coming out of 24XStream an going into ATmega28 Figure 4. Communication latency of a single wireless transmission. 30

31 [ ] G = 0 ; m q s = ; for i =,2, Solve equation (a) for Solve equation (b) for H i ; L ; ( ) i Fin graient using equation (c): 2 T ( ) T 2 T i = + + iterate { G = + i+ i s G i ; Solve equation (a) again for i+ if trace ( Hi+ Z l) < trace( i l) max B H A B G D LD RG D LD ; H using G i+ ; HZ an ( eigen ( A + BGi+ D ) ) < exit the iterate loop; else s = s / 2; If (s < machine precision), then exit the iterate loop; en }; s = s 2; If G i+ G i < acceptable error, then exit the for loop; en Figure. 5. Heuristic algorithm solving the couple nonlinear matrix equations (Eq. ) for centralize optimal time-elay output feeback control (Lunze, 990). 3

32 Maximum Drift Among Three Stories LQR Inex Performance Inex PI Sampling Time (s) 0 (a) 2 3 Degree of Centralization Performance Inex PI Sampling Time (s) 0 (b) 2 3 Degree of Centralization Performance Inex PI Maximum Drift Among Three Stories - Decentr. 2 - Partially Decentr. 3 - Centr. Performance Inex PI LQR Inex - Decentr. 2 - Partially Decentr. 3 - Centr Sampling Time (s) (c) Sampling Time (s) () Figure 6. Simulation results illustrating control performance inexes for ifferent sampling time steps an centralization egrees: (a) 3D plot for performance inex PI ; (b) 3D plot for performance inex PI 2 ; (c) conense 2D plot for PI ; () conense 2D plot for PI 2. 32

33 (b) (a) (c) Figure 7. Laboratory setup: (a) the 3-story test structure mounte on the shake table; (b) the MR amper installe between the st floor an the base floor of the structure; (c) a wireless control unit an an off-boar control signal generation moule. 33