Health Monitoring of Civil Infrastructures Using Wireless Sensor Networks

Health Monitoring of Civil Infrastructures Using Wireless Sensor Networks Sukun Kimt, Shamim Pakzadt, David Cullert, James Demmelt Gregory Fenvest, Steven Glasert, Martin Turon* {binetude, culler, demmel}@eecs.berkeley.edu {shamimp, fenves, glaser}@ce.berkeley.edu mturon@xbow.com tElectrical Engineering and Computer Sciences and tCivil and Environmental Engineering *Crossbow Technology, Inc. University of California at Berkeley 4145 N. First Street Berkeley, CA 94720 San Jose, CA 95134 ABSTRACT etc.). The general approaches taken to SHM are either direct dam- A Wireless Sensor Network (WSN) for Structural Health Moni- age detection (visual inspection, x-ray, etc.) or indirect damage de- toring (SHM) is designed, implemented, deployed and tested on tection (detecting changes in structural properties or system behav- the 4200ft long main span and the south tower of the Golden Gate ior). This paper describes a platform for indirect detection of struc- Bridge (GGB). Ambient structural vibrations are reliably measured tural state through the measurement and interpretation of ambient at a low cost and without interfering with the operation of the bridge, vibrations and strong motion. The chosen test bed is the Golden Requirements that SHM imposes on WSN are identified and new Gate Bridge in San Francisco Bay. Performing SHM by the use of solutions to meet these requirements are proposed and implemented. sensor networks is not a new concept [20, 9]. The traditional ap- In the GGB deployment, 64 nodes are distributed over the main proach consists of conventional piezoelectric accelerometers hard- ' ... . . . ~~~~wired to data acquisition boards residing in a PC. The drawbacks span and the tower, collecting ambient vibrations synchronously at qug IkHz rate, with less than 10,s jitter, and with an accuracy of 30,uG. of such a system include (1) the high cost of installation and distur- The sampled data is collected reliably over a 46-hop network, with bance of the normal operation of the structure due to wires having a bandwidth of 441B/s at the 46th hop. The collected data agrees to run all over the structure, (2) the high cost of equipment; and with theoretical models and previous studies of the bridge. The (3) cost of maintenance. Compared to the conventional methods, deployment is the largest WSN for SHM. Wireless Sensor Networks (WSN) provide comparable functional-ity at a much lower price, which permits a higher spatial density Categoriesand Subject Descriptors of sensors. The prototype wireless system presented in this paper costs about $600 per node compared to thousands of dollars for a C.3 [Special-Purpose and Application-Based Systems]: Real- node with the same functionalities in a traditional PC-based wired time and Embedded Systems network. Compared to the wired network, installation and main- tenance are easy and inexpensive in a WSN, and disruption of the General Terms operation of the structure is minimal. Experimentation, Reliability, Design This work has three main contributions to WSN for SHM: * It identifies requirements to obtain data of sufficient quality Keywords to have real scientific value to civil engineering researchers, Wireless Sensor Networks, Structural Health Monitoring, Deploy- and examines how to solve them. ment, Large-Scale * The system is scalable to a large number of nodes to allow dense sensor coverage of real-world structures. For example,1. INTRODUCTION a long-lived 46-hop network was implemented on the Golden Structural Health Monitoring (SHM) is a technology that allows Gate Bridge (Figure 1 and 2). the estimation of the structural state and detection of structural change that affects the performance of a structure. Two discriminat- * It addresses a myriad of problems encountered in a real de- ing factors in SHM are the time-scale of the change (how quickly ployment in difficult conditions, rather than a simulation or the state changes) and the severity of the change. These factors rep- laboratory test bed. resent two major sources of system change: alarm warnings [27] A WSN for SHM was deployed on the Golden Gate Bridge (GGB),(e.g. disaster notification for earthquake, explosion, etc.) and con- ' ' ~~~~~~seeFigure 1. The 46-hop system consists of 64 nodes, which mea-tinuous health monitoring (e.g. from ambient vibrations, wind, sure ambient vibrations with an accuracy of 3OjiG. The ambient vi- brations were sampled at 1kHz with a time aperture less than 10,s. Figure 2 illustrates the bandwidth obtained by Straw, a new reliable Permission to make digital or hard copies of all or part of this work for data collection component written for this installation. The system personal or classroom use is granted without fee provided that copies are provided high bandwidth data streaming of 441B/s from the 46th not made or distributed for profit or commercial advantage and that copies hop to the base station by implementing pipelining. This 46-hop bear this notice and the full citation on the first page. To copy otherwise, to wireless sensor network is the largest number-of-hop installation republish, to post on servers or to redistribute to lists, requires prior specific reported in the literature up to now. permission and/or a fee. IPSN'07, April 25-27, 2007, Cambridge, Massachusetts, USA. This paper will explain how the system was designed and imple- Copyright 2007 ACM 978-1-59593-638-7/07/0004 ...... ...........$5.00.mented to achieve this successful deployment on the bridge. The 254 Bandwidth versus Hop CountSF Sausalito (south) ft north) Aug 1 st_------------------------------------------------X1200 ~~ '¢('_ : J', YJLvUBU@B@U@U@UBU@UBUBUB@@@@@@@ J X \\ ~~~~~~~~~~~~~~~~~~~Sep20th -----125ft- 2A-0.4200ft 100----IUU ...... ..................................... ..... ........... 56 nodes 100 S 2 ~- 800 8 nodes 600 , Figure 1: The Golden Gate Bridge and layout of nodes on the C5 _X'X bridge. To cover this large bridge, long linear topology needs n 400 be used, and it brings challenges to the network. 200 0 later part of the paper will present an analysis of the collected ac- 0 5 10 15 20 25 30 35 40 45 50 celeration data recorded along a linearly dense array. Six major Hop Count requirements of SHM on WSN are identified here. 1. The data acquisition system had to be able to detect signals Figure 2: Bandwidth of Straw at the Golden Gate Bridge. It with peak amplitudes as low as 500,uG [7]. The installa- works over a 46-hop network. To sustain high bandwidth over tion had to minimize sources of distortion such as the noise a long path, pipelining is used avoiding interference. floor of the system (including accelerometer, amplifier, ana- log to digital converter, etc.), installation error, and tempera- ture variation. Tenet [12] satisfy many requirements. They provide reliable data collection over multi-hop network. However, Wisden can sample 2. Because of structural interest in local modes of vibration, a only up to 160Hz, and Tenet demonstrated only 50Hz sampling, sampling rate of 1kHz was chosen as the target rate. This which is far below the threshold needed for structural health mon- rate and the need for 16-bit digitization accuracy require low itoring. Wisden and Tenet have not been analyzed for sampling jitter, i.e. low time uncertainty of the sampling intervals. jitter, which is needed in determining to what degree the result- 3. Time synchronization in sampling through the bridge is re- ing data has confidence for analysis in civil engineering. They are quired to perform correlation analyses of the structural vibra- tested only in a small-scale indoor test bed. The most critical pitfall tions. This was particularly challenging due to the drift of in- of these two systems is that they do not produce time-synchronized dependent clocks at each of the 64 nodes. An earlier reported data. Wisden has a time stamp on each sample. However, the input solution to the time synchronization problem is FTSP [19]. for basic modal property analysis is a matrix of time-synchronized samples from multiple nodes. The data produced by Wisden and 4. The GGB installation required a large-scale multi-hop net- Tenet has no value for meaningful structural analysis. work due to the great length of the main span and the fact that the aggregator station could only be located in the base 3. OVERALL ARCHITECTURE of the south tower. One existing solution to the collection networkisohMntRoue [26].existing solution tothecollection The wireless network is composed of multiple nodes and a base station. A node consists of a mote and a sensor board. The node 5. Commands had to reliably disseminate throughout the entire measures vibration at two different orders of dynamic bandwidth, system so that all parts of the network could start on com- with the data communicated back to the base station through wire- mand, and insure against lost data or a blockage of hopping. less communication. The base station is a server providing more Repeated Broadcast [2] can be one solution. computational power and larger storage than a mote node, and pos- sibly a connection to the Internet. In the GGB deployment a laptop6. Data must be transferred reliably. Vibration data, in this case, isuds base tation. th sh the GGB ' ' 1~~~Sused as a base station. The software architecture of the GGBis too valuable to be lost to communication error. nodes uses new components integrated into the TinyOS [13] infras- tructure to satisfy the six requirements discussed above. Figure 3 2. RELATED WORK illustrated the overall software structure. A low latency dissemi- WSN applications can be divided into two categories. The first nation service is required so that the commands are not delivered category is environmental monitoring; networks deployed in Great after they are needed, therefore Broadcast [2] was used in place of Duck Island [22] and a Redwood forest [24] are examples of this Drip [23]. Drip provides dissemination service with an eventual re- class. For this class of problem the focus is on networks with low liability but has long latency. Even though Broadcast provides un- duty-cycle and low power consumption. The second category of reliable dissemination service, with repeated broadcast 100% even- WSN applications consists of applications that require identifica- tion of a mechanical system through a measured system response. Health monitoring of mechanical machines [16], condition-based monitoring, volcano monitoring [25], earthquake monitoring [11], _ and structural health monitoring [21] belong to this class, which S FTaw Bfgenerally require high fidelity sampling. The focus of this paper Broadcast NMintRout is to address the requirements of the latter category. Related work 1 on using WSN in SHM includes [10, 8, 14, 18] . However, these Best-ffor Single-hop Communication ow-level FLAS1 networks generally do not scale to a long enough multi-hop net- work needed to cover a large structure, and have not been imple- mented and tested in a harsh real-life environment. Wisden [27] and Figure 3: Overall Software Architecture 255 Thermometer ADXL 202E Accelerometer Board Mote S ~~M~~~ll M~~~J\ ~~Raidiod SD1221 -assFL SDI22IL -assF D : FlashSilicon Designs 1221L Figure 5: Accelerometer Board. ADXL 202E has two axis in a Figure 4: Hardware Block Diagram. Details of two accelerome- single chip. Either Mica2 or MicaZ can be used as a mote. ters (ADXL 202E and SD 1221L) are in Table 1. A thermometer is used for temperature calibration. Table 1: Comparison of the Two Accelerometers. G means the acceleration of gravity. tual reliability can be achieved in practice. MintRoute [26] was ADXL Silicon Designs used for information reply since it provides a best-effort multi-hop 202E 1221L convergence routing. Our new reliable data collection layer Straw Type MEMS MEMS lies above Broadcast and MintRoute. For time synchronization, Range of System -2G to 2G -0.1G to 0.1G FTSP [19] is used. BufferedLog [3] supports high frequency sam- System noise floor 200(,uG /Hz) 32(j(G/VHfz) pling with light-weight logging. Structural hEalth moNiToRing toolkIt (Sentri) is an application Price $10 $150 layer program which drives all components. Instead of a stand- alone program, Sentri is structured like an RPC server: for every operation a command is sent from the base station to a node. In loading and traffic, are resolved by a two-dimensional SiliconDe- SHM, motes are heavily used and heavy traffic makes network signs 1221L accelerometer. A low-cost ADXL202E two-dimensional bandwidth the bottleneck of the operation; therefore, additional accelerometer was used to monitor stronger shaking as might be processing and traffic overhead must be avoided. However, since expected from earthquake excitation. Because input battery power the project is in the research stage in both system engineering and can vary between 6V and 12V, the sensor board contains a voltage civil engineering and the operation sequences and parameters were regulator to provide a constant 3V output for the mote and a con- changing frequently, the operational model was necessary. It al- stant 5V output for the ratiometric accelerometers. Table 1 presents lows us to figure out precisely which parts of the signals are more the characteristics of the two accelerometers used, and associated valuable, and to fine-tune the system parameters in an interactive analog circuits. Two simple filters are used on the board. One is a process. This process was first tested and verified through a trial hardware-implemented single-pole 6db low-pass filter with a cut- deployment of several nodes on a model steel bridge at UC Berke- off frequency of 25Hz. Since the on-board ADC quantizes much ley and then again on a footbridge over 1-80 [21], both of which faster than the target sampling frequency, this extra capacity allows were used to determine system settings. Sentri provides 16 opera- on-the-fly digital filtering after a factor of Sover = 10 oversam- tions and the command for each operation is contained in a single pling, and then averaging the samples before logging. Assuming packet. Operations like reset, erase flash, start sensing, and reading a Gaussian distribution for the noise, oversampling by a factor of meta-data are examples of such high-level Sentri commands, which Sover = 10 reduces the noise level by a factor of Sover -_ 3.16. provide a great deal of flexibility and can be sent in sequence by the Another key hardware consideration in WSN is power consump- base station. tion. The high duty-cycle required by vibration SHM produces data sets that are between two to four orders of magnitude larger 4. DATA ACQUISITION SYSTEM than that of an environmental monitoring application. In contrast4. DATA ACQUISITION SYSTEM to environmental monitoring, this application requires continuous Figure 4 shows an overview of the hardware as a block diagram. operation. The gain from duty-cycling does not provide compelling The data acquisition system performs three primary functions: sens- savings compared to its complexity and overhead, so duty-cycling ing, signal processing and communication. Because of long ex- is not used. The higher data volume requires sophisticated on- perience with the product, Crossbow MicaZ [5] motes were used board computation with a distributed system identification algo- for control and communications. The analog signals output by the rithm (which is expensive in terms of energy), or all the data needs low-noise accelerometers pass through low-pass antialiasing filters to be transmitted to a base station for further processing (which is on the way to a 16-bit analog-to-digital converter, before the data is even more power-expensive). The use of batteries or other renew- first logged into the flash of the mote and then wirelessly transmit- able sources of energy is justified for quick and temporary appli- ted. cations, or where a more permanent power source cannot be pro- vided. An analysis of the power consumption of the boards was 4.1 ACCELEROMETER SENSORBOARD performed to determine the size of the batteries. In the deploy- A new accelerometer board [5], shown in Figure 5, was designed ment at the Golden Gate Bridge, 4 lantern batteries are used for for SHM applications. The board has four independent accelerom- each node. Table 2 shows the actual power consumption profile of eter channels monitoring two directions (vertical and transverse), a complete sensor unit, from which it is seen that the sensor board and a thermometer to measure accelerometer temperature for com- by itself consumes about twice the energy of the mote. The board pensation purposes. Low-amplitude ambient vibrations, due to wind design had a single power path for the mote, sensors, and ADC. 256 Table 2: Power Consumption in Various Operational Situations (9V input voltage). Idle is when both the sensor board and the Node 1 \Spatial jitter mote are turned on, but are not performing any operation. Situation Consumption (mW) Node 2 Temporal jitter Board Only 240.3 Mote Only 117.9 Node 3 Time Idle 358.2 Node3____ Time One LED On 383.4 Erasing Flash 672.3 Figure 6: Sources of Jitter. Both temporal jitter and spatial Sampling 358.2 jitter should be within a threshold for the data to have scientific Transferring Data 388.8 value. Significantly lower energy consumption could be realized if only component declares them to be in sync. For a target-sampling rate the mote is directly connected to the battery, so that all other com- of 200Hz, a total jitter of 250,us or 5% of the sampling interval was ponents can be turned off when the unit is not collecting data. selected as the cap to total jitter. A study of the time synchroniza- 4.2 CALIBRATION tion component FTSP showed that it caps jitter at 67,us over a fifty- nine node eleven-hop network [19], so spatial jitter in this case is The static noise floor of our accelerometer devices was quantified within the tolerance range. Temporal jitter can become larger than in the Berkeley Seismological Laboratory underground seismome- spatial jitter during periods of high-speed data collection, so this ter calibration vault. Testing showed that the SiliconDesigns 1221L was studied in detail. In particular, we explored and modeled tem- devices have a noise floor of poral jitter, and show that our model matches measured data. We 321,GI /Hz, which is small enough to allow resolution of the am- will also show that jitter cannot be completely removed without bient vibrations of most structural systems. Examples of similar adding another microcontroller. measurement systems in civil infrastructures can be found in [7]. Shaking table tests with patterns ranging from 0.5Hz to 8Hz were 5.1 TEMPORAL JITTER ANALYSIS performed to study the dynamic behavior of the accelerometers. A statistical model may not catch every minute detail of the tem- The accelerometers perform well within the expected dynamic range .[21].heachaccelerometers channrmwell washinra pe-ca ted umicangeaporal jitter process, but it will provide understanding of the dis-[21]. Each accelerometer channel was range-calibrated using a tilt trbto.ftmoa jte.Th'ie vetfrsmligtcsatt tribution of temporal jitter. The timer event for sampling ticks at test process. The boards were attached to a tilting machine [6], uniform intervals is graphically presented in the upper portion of which has a rotational accuracy of 0.001 degree, and the digital output correlated to each angle tested. All four channels showed ofigurei. hngother timer,event fires,ttheCPUacanobeRin the middlOf servicing other tasks, such as writing data from RAM to flash.linear response, and the testing provided offset and scale factors. When the CPU is servicing an atomic section, the timer event is Prototype accelerometers were also tested in an oven to study the response of the devices to temperature. The tests showed that not dela