a localized algorithm for structural health monitoring using wireless

8/13/2019 A Localized Algorithm for Structural Health Monitoring Using Wireless

1/16

A localized algorithm for Structural Health Monitoring using wirelesssensor networksIgor Leo dos Santosa, , Luci Pirmeza, rico T. Lemosa, Flvia C. Delicatob, Luiz A. Vaz Pintoc, J. Neuman de Souzad, Albert Y. Zomayaea iNCE Universidade Federal do Rio de Janeiro (UFRJ), CEP 21941-916, Cidade Universitria, Rio de Janeiro, RJ, Brazilb DCC, Universidade Federal do Rio de Janeiro (UFRJ), CEP 21941-916, Cidade Universitria, Rio de Janeiro, RJ, Brazilc Departamento de Engenharia Naval e Ocenica, Universidade Federal do Rio de Janeiro (UFRJ), CEP 21941-909, Cidade Universitria, Rio de Janeiro, RJ, Brazild Universidade Federal do Cear (UFC), CEP 60455-760, Campus do Pici, Fortaleza, CE, Brazile Advanced Networks Research Group, School of Information Technologies, The University of Sydney, NSW 2006, Australia

a r t i c l e i n f o

Article history:Received 25 April 2010Received in revised form 2 February 2012Accepted 7 February 2012Available online 16 February 2012

Keywords:Wireless sensor networksStructural Health MonitoringDamage localizationLocalized algorithmInformation fusionResource constrained networks

a b s t r a c t

Structural Health Monitoring (SHM) has been proving to be a suitable application domain for wirelesssensor networks, whose techniques attempt to autonomously evaluate the integrity of structures, occa-sionally aiming at detecting and localizing damage. In this paper, we propose a localized algorithm sup-ported by multilevel information fusion techniques to enable detection, localization and extentdetermination of damage sites using theresource constrained environment of a wireless sensor network.Each node partakes in different network tasks and has a localized view of the whole situation, so collab-orationmechanismsandmultilevel informationfusion techniques arekeycomponentsof thisproposal toefciently achieve its goal. Experimental results with the MICAz mote platform showed that the algo-rithm performs well in terms of network resources utilization.

2012 Elsevier B.V. All rights reserved.

1. Introduction

Recently, there has been much interest in the use of WSNs [1] inthe eldsof exploration anddistributionof the oil andgas industryas well as in the renewable energy sector, particularly in windfarms, with the purpose of Structural Health Monitoring (SHM)[2]. The monitoring of physical structures enables damage predic-tion (fractures) and, therefore, repairs anticipation thus avoidingaccidents. In applications built for that purpose, the sensor nodesare used to perform measurements of the structure which isaffected by external events, delivering such measures to a data col-lection station, the sink node. In this context, WSNs enable the re-mote monitoring of structures to determine physical integritythrough in situ data collection and processing .

This work proposes a localized algorithm, calledSensor-SHM, todetect, localize and indicate the extent of damage on structuresbelonging to environments like offshore oil and gas industry andwind farms, making use of WSNs for a SHM system. The topology

of the WSN is assumed to be hierarchical, where sensors aregrouped into clusters and each cluster is managed by a cluster-head (CH). The key idea of our work is to fully distribute the pro-cedure associated with the task of monitoring a structure amongthe sensor nodes in a WSN, so that through collaboration amongthe CHs it is possible to detect, localize and determine the extentof damage. Unlike other approaches [3,4], all the SHM processingof our proposal runs inside the network (in-network processing)without any help from the sink node. When distributing the SHMprocessing inside the network, our work strongly takes advantageof using information fusion techniques, whose immediate benetis the reduction in the amount of data to be transmitted back tothe sink node for further analysis. Consequently, less energy isspent due to transmissions, enabling the use of communicationand energy resources for performing analysis and taking decisionswithin the network. In our proposal, we make use of a terminologyreviewed by Nakamura et al. [5] regarding informationfusion tech-niques applied to WSNs. Such terminology was originallyproposedby Dasarathy [6] in its DataFeatureDecision (DFD) model. Theterminology points three different abstraction levels of the manip-ulated data during the fusion process: measurement , feature anddecision . The whole process considered in our proposed algorithmcan be classied as a Multilevel Fusion , since it acts in the threeexistent data abstraction levels.

1566-2535/$ - see front matter 2012 Elsevier B.V. All rights reserved.doi:10.1016/j.inffus.2012.02.002

Corresponding author.E-mail addresses: [email protected] (I.L. dos Santos), luci.pirmez@gmail.

com (L. Pirmez), [email protected] (.T. Lemos), [email protected] (F.C.Delicato), [email protected] (L.A. Vaz Pinto), [email protected] (J.N.de Souza), [email protected] (A.Y. Zomaya).

Information Fusion 15 (2014) 114129

Contents lists available at SciVerse ScienceDirect

Information Fusion

j ou rna l homepage : www.e l sev i e r. com/ loca t e / i n ffu s
http://dx.doi.org/10.1016/j.inffus.2012.02.002mailto:[email protected]:luci.pirmez@gmail.%20commailto:luci.pirmez@gmail.%20commailto:[email protected]:[email protected]:[email protected]:[email protected]:[email protected]://dx.doi.org/10.1016/j.inffus.2012.02.002http://www.sciencedirect.com/science/journal/15662535http://www.elsevier.com/locate/inffushttp://www.elsevier.com/locate/inffushttp://www.sciencedirect.com/science/journal/15662535http://dx.doi.org/10.1016/j.inffus.2012.02.002mailto:[email protected]:[email protected]:[email protected]:[email protected]:[email protected]:luci.pirmez@gmail.%20commailto:luci.pirmez@gmail.%20commailto:[email protected]://dx.doi.org/10.1016/j.inffus.2012.02.002


2/16

This work builds on our previous work [7], introducing severalenhancements on it. The main contributions of our previous workare: (i) we introduced the core of the algorithm with no specicfoundation, which could help to better understand the algorithmoperation, and (ii) several experiments to evaluate the consump-tion of energy resources from the network were performed. Weaugmented our previous work with the main following contribu-

tions, discussed in this paper: (i) we provided a foundation forthe proposed algorithm that relies on the information fusion the-ory, (ii) several new experiments were performed concerning theprecision of our damage localization mechanism, and (iii) a com-prehensive analysis on the use of communication resources inthe algorithm was provided.

The remainder of this paper is divided as follows. Section 2 pre-sents an overview of Sensor-SHM algorithm. Section 3 depictsrelated works. Section 4 presents a motivational example basedin thepracticalexperiment seen in [8] anddiscusses the applicabil-ity of the algorithm. Section 5 presents the algorithm, discussingand detailing its procedures. Section 6 details the experiments per-formed to evaluate Sensor-SHM algorithm and the obtainedresults. Finally, Section 7 concludes this work.

2. Overview of the algorithm

The diagram in Fig. 1 presents an overview of the proposedalgorithm, Sensor-SHM, and its procedures. In this diagram, theroles of sensors and CH nodes are summarized, and the setup pro-cedure and data collection stages are presented separately.

After the execution of a setup procedure (Procedures 04), eachsensor node acts in the data abstraction level named measurement,delivering the rst useful features to its respective CH. Each sensoris responsible for sensing the structure during a data collectionstage (Procedures 516), which starts from the sink node throughthe transmissionof messagesto theCHs, whichrequest thesensingof the structure by the sensors in their respective clusters (Proce-dures 5 and 6). Then, each sensor node acts collecting the acceler-ation measurements in the time domain, relative to its physicalposition (Procedure 7). After that, a Fast Fourier Transform (FFT)is performed by each sensor over the collected acceleration signals(Procedure 8). Such transformation corresponds to the informationfusion technique classied as a Data InData Out (DAIDAO). Next,a method for extracting frequency values from the peaks of thepower spectrum generated by the FFT is used (Procedure 9), whichcan be composed of a moving average lter (another example of DAI-DAO information fusion technique, where the input is thepower spectrum, and the output is the smoothed power spectrum)and the peak extraction algorithm itself, applied on the smoothed

power spectrum. This peak extraction algorithm is an informationfusion technique classied as a Data InFeature Out (DAIFEO),since the extracted peaks are the rst features which are consid-ered useful to describe the structural health state and can beefciently manipulated among the sensors. The frequency valuesobtained in each sensor refer to the rst peaks of the power spec-trum returned by the FFT, and will make up the signature of the

structure. It is important to mention that for each sensor the initialsignature of the structure is obtained from its current position inthe beginning of the structure operation, i.e., at time zero, and istransmitted, during the network setup procedure by the sensorto its CH (Procedure 3). This signature is used as a reference forthe undamaged structure. At later stages, each CH also receivesthe subsequent signature of the structure of all the sensors in itsrespective cluster (Procedure 10).

CHs are responsible for performing the damage detection anddetermining the damage location and extent through the calcula-tion and analysis of damage coefcients (Procedures 1114). TheCH, after collecting the signatures from all sensor nodes of its clus-ter, performs a comparison (considering a given tolerance degree)betweenthese values andthe respective initial signatures from therespective sensors, to check whether the structure is damaged or ithas been temporarily changed due to some external event. At thispoint, the CH starts its own sequence of information fusionprocedures, acting in two levels of data abstraction ( feature anddecision ).

The presence of damage on a structure can affect both higherand lower frequencies, in a given sensor location, depending,respectively, if the sensor is located close to the damage or not[3]. Knowing that changes in the frequencies of the higher vibra-tion modes mean changes in local vibration modes, each CH ana-lyzes the signatures of the sensors located in its cluster in searchof changes in these frequencies. In each CH and for each data col-lection stage, this analysis is performed with the help of a damageindicator coefcient ( Di,t ). The value of Di,t indicates how close agiven sensor location is to the damage site. This rst damage coef-cient is the result of applying a Feature InFeature Out (FEIFEO)information fusion technique. A second damage indicator coef-cient for the cluster ( C j,t ), which depends on the Di,t coefcientsobtained for each sensor from the cluster, is set to indicate howclose to the damage the cluster is as a whole. In this last technique,a Feature InDecision Out (FEIDEO) information fusion technique,each CH node compares its cluster damage coefcient with a toler-ance, which is differently set for each CH depending on the specicfeatures of the places where the cluster was installed. When thecluster damage coefcient exceeds the tolerance, the CH nodeshould send a message stating the value of its coefcient to itsimmediate (single-hop) neighbor CHs (Procedures 13 and 14). In

Fig. 1. Overview of the proposed algorithm and its procedures.

I.L. dos Santos et al. / Information Fusion 15 (2014) 114129 115


3/16

a given neighborhood, the CH with the highest value of clusterdamage coefcient assumes the role of a collector responsiblefor (i) issuing a warning and (ii) triggering a relay acting on theenvironment around it, aiming to prevent the damage progressionto avoid causing further problems in the locality (Procedures 15and 16). When using the highest cluster damage coefcient amongthe CHs, an aggregation method based on a maximum aggregation

function is being performed to dene the collector. The collectorwill collect the decisions of each CH in its neighborhood to makeits nal decision. So this nal collaboration among CHs canbe clas-sied as a Decision InDecision Out (DEIDEO) information fusiontechnique, since each CH has its own local decision for assumingdamage or not, but it is allowed to consider the decisions of otherlocal neighbors to present a more reliable decision. The joint use of all the processes mentioned in our proposed algorithm, each oneexisting in one or more levels of information fusion, is thereforethe reason to classify it as a Multilevel Fusion .

3. Related work

The related works presented in this section are classied basedon the concept of generations of WSNs for SHM,which is an expan-sion to the concept of a rst-generation wireless structural mon-itoring system, used in [9]. We state that the rst generation of sensor networks for SHM is relative to wired devices, while the sec-ond generation of sensor networks for SHM is related to wirelessdevices. The second generation is further divided in two groups:Centralized Generation of WSNs for SHM and DecentralizedGeneration of WSNs for SHM . These two groups differ with respectto the degree of decentralization and in-network processing pre-sented by their works.

Centralized Generation of WSNs for SHM is characterized bycentralized proposals, with few in-network processing aimed atdamage characterization.Most of the in-network processing occursin the measurement level of information fusion, for reliable andefcient raw data transport. Works which t this classicationare for instance [3,810]. Decentralized Generation of WSNs for SHM is characterized by some degree of decentralization in itsworks. Some of these works aim at the calculation of damageindexes which can be efciently transmitted by the wireless net-work and the calculation of the damage indexes is performed overthe rawaccelerationdata. Also, fewlevels of information fusionarepresented by these works. Our work pertains to this generation asa fully decentralized proposal.

Caffrey et al. [11] present an algorithmfor SHM using WSNs andthis algorithm is classied by us into the second generation of WSNs for SHM. The sensing technique in the related work useselectrodynamic shakers to generate vibrations on the structure,and uses accelerometers in the sensor nodes to collect data for a

few seconds, in order to capture these vibrations. In order to per-forma structural analysis, a Fast Fourier Transform (FFT) is appliedover the acceleration data collected in each sensor node, convert-ing the signal in time domain to a signal in frequency domain.Then, the power spectrum is analyzed and the frequencies of thestructural modes of vibration, whose values correspond to theenergy peaks of the spectrum, are extracted. Then, onesensor nodeis elected among all the sensors to be responsible for obtaining amore accurate result by aggregating all measures of modal fre-quencies and their associated energies extracted from all the net-work nodes. Finally, this aggregated result is sent to the sink andthen the frequency variation analysis can be done. Our proposaldiffers from this one since the frequency variation analysis is per-formed within the network, with the collaboration of CHs.

Chintalapudi et al. [4] present three methods to detect damageon structures. The rst one is a technique that performs data

reduction within the network, and is based on a time series of acceleration signals. In this method, the response of a structure ismodeledusing linear auto-regressive (AR), or auto-regressive mov-ing average (ARMA) time series. The damage is detected by a sig-nicant variation in the AR/ARMA coefcients, relative to theintact (healthy) structure coefcients. Each sensor node can locallycompute these coefcients and forward them, instead of requiring

that all sensors transmit their collected acceleration data. The sec-ond method consists in the methodology for damage detection instructures that makes use of the structures signature variation,and it is widely accepted. This variation is observed when signa-tures obtained when the structure is sound and signaturesobtained when the structure is damaged are compared. The practi-cal results of this methodology are found in [12,13]. The lattermethod makes use of a neural network to detect the possibilityof damage. Most studies found in the literature are inspired inthe second method, including our proposal. Unlike Messina et al.[13], in which analytical values of the structure natural frequenciestaken from a niteelement model areused, thesystemproposed inour work uses frequency values taken when the structure is sound.Another difference is that our algorithm runs on the sensors andthe algorithm called Damage Location Assurance Criterion (DLAC)by Messina et al. [13], runs on the sink node.

Hackmann et al. [14] make use of a WSN to monitor structuralconditions. The proposed partially decentralized algorithm is usedalong with the DLAC method, allowing the sensor nodes to act onthe collected data, signicantly reducing energy consumptionsince it minimizes the number of transmissions needed. In thealgorithm, the data is collected and partially in-network processedby the sensor nodes. In the sink node, the frequency values thatcompose the signature of the structure are extracted by solving amathematical equation expressing a curve that ts the resultantpower spectrum. After, theDLAC algorithmruns on the sink, takingdata relative to two sources as input to detect and locate damage:(i) the sensed data relative to the structural frequency responseand (ii) data relative to the responses of an analytical model tothe same scenario. The analytical model is developed through anite element modeling. It is important to note that the damagedetection and localization through the use of the DLAC methodwas still centrally held, in the sink node. The algorithm for damagedetection proposed in our work is mainly inspired by thework pre-sented in [14]. One of the differences between our proposal andsuch work is that in our solution the whole procedure of extractingthe frequency values from the power spectrumis performed on thesensors, while in the related work it is conducted by the Curve Fit-ting stage. Unlike other algorithms proposed in the literature, allthe SHM processing of our algorithm is performed on the sensorand CH nodes, without the help of the sink node. And through col-laboration among CH nodes, it is possible detect, localize, anddetermine the extent of damage.

It is also noticeable, in the literature, the widespread use of damage coefcients. In general, the numerical values of these coef-cients are extracted from raw data through information fusiontechniques, and their values indicate how intense a damage occur-rence is, if it exists, or how close a site is to the damage site. Wanget al. [15] analyze the design requirements of WSN-based SHMapplications and discuss the related challenges dealt with in dis-tributed processing, presenting two algorithms. The rst algo-rithm, distributed damage index detection (DDID), is a distributedversion of a centralized algorithmpreviously proposedby the sameauthors. The second algorithm, collaborative damage event detection(CDED), uses the results provided by the DDID algorithm and aimsto improve the reliability and accuracy of the damage report byexchanging data among the sensor nodes.

By using the DDID algorithm, every sensor node inspects theraw data and determines the damage candidates. Once the

116 I.L. dos Santos et al. / Information Fusion 15 (2014) 114129


4/16

candidates are captured, the sensor nodes in a group will use theCDED algorithm to cooperatively create the damage report con-taining the information about the location, scale, and index of the damage. CDED and DDID are complementary algorithms inwhich information fusion is performed in different abstraction lev-els. In DDID, the measurement and feature levels are more evident,whilst in CDED the decision level is more evident, what character-

izes a multilevel data fusion as a whole. The sensors are clusteredin this related work, and the network is comprised of two tiers, onecontaining the sensor nodes and a second one containing the mas-ter nodes. Moreover, two main issues are pointed: (i) how muchinformation should still be sent back to the server in order to en-sure the energy efciency and results precision, and (ii) how toautomate the decision on the existence, localization, and extentof damages which is originally made through human interference.

Our proposal is also based on a multilevel information fusiontechnique to generate damage indexes, and we have also adopteda hierarchicalnetwork organization. Both issues (i) and (ii) are alsodealt with in our work. So, the main points in which our work dif-ferentiates from the proposal of [15] are (i) the way in which thedamage coefcient is calculated, since our damage coefcient isbased on the analysis of modal frequencies and this feature isnot exploited in DDID, which makes use of raw acceleration sig-nals, (ii) our algorithm proposes more operations of informationfusion in different levels to calculate the damage coefcients and(iii) in our proposal the cluster meaning is closely related to SHMissues and its behavior is different.

The centralized proposals are generally represented by wirednetworks with no in-network processing for assessing the struc-tural integrity. The common approach of such solutions is com-pletely opposite to the concept of network-centric solutions, andthe presence of wires forbids large-scale deployments due to tech-nical and economical issues. Both generations which are repre-sented by WSNs present increasing levels of data compression,aggregation and fusion, with more distributed processing aimedat physical integrity characterization (of the structures). Therefore,the reason to use a decentralized approach is to achieve a longerlifetime for a wireless system which can monitor a structure moreexibly than a wired system. The reduction in the need for trans-missions in a decentralized approach is the main cause of thereduction in energy consumption. Nevertheless, the decentralizedapproach also points towards new challenges. The resource con-strained environment of the WSNs forbids the extensive use of computational resources. The lack of memory and processinggenerally imposes restrictions and trade-offs which may affectnegatively the precision and accuracy of the decentralized pro-posal. Also, the energy constraints cause the need to schedule theamount of sampling tasks, or the frequency of the monitoring cy-cles, very carefully.

4. Motivational example and the applicability of the algorithm

In order to illustrate theapplicability of the proposedalgorithm,an example of a real application described in Clayton et al. [8] waschosen. This application is performed over a test structure in a con-trolled laboratory environment, what corresponds to the majorityof the application scenarios adopted in the literature. The experi-ment described in the work of Clayton et al. was performed to val-idate the performance of MICAz motes for damage detection andlocalization. The considered test structure consisted in a ve baylumped mass shear building model. MICAz motes were installedon oors 1, 2, 4 and 5, and recorded acceleration in the x-axisdirection. Damage was simulated through the reduction of the

inter-storey column stiffness, by exchanging the original columnsby those with a lesser moment of inertia. Thestructurewas excited

by an impulse load generated by striking the fourth oor with amodally tuned impact hammer. The results were also recordedby wired sensors, and compared to the results of the MICAz plat-form, showing that the WSN platform performed well in compari-son to the wired nodes. In the experimental scenario, the moteswere at less than 1 m distance from each other, in average. Thealgorithm proposed in our work is applicable to this scenario, as

well as to many others. After presenting several evaluations onthe performance of our algorithm, in Section 6.8 we present anddetail a procedure to deploy a wireless sensor network runningSensor-SHM. The procedure is based on the real experiment de-scribed in [8]. The scenario parameters were used to simulate thebehavior of our algorithm analytically.

5. The proposed localized algorithm: Sensor-SHM

Since Sensor-SHM is a distributedalgorithmin whicheach nodeof the network has only a partial view of the global situation andnodes collaborate by sharing their views to achieve the nal goal,it can be classied as a localized algorithm [16]. Moreover, Sen-sor-SHM can be classied as a multilevel information fusion algo-

rithm, since it encompasses a sequence of information fusionprocedures, each of which acting in one or more of the three dataabstraction levels. The description of Sensor-SHM is divided intotwo main procedures. Initially, a setup procedure is performed,which consists in setting the algorithm initial parameters beforemaking the application and network deployment, i.e., beforeinstalling the program on the sensors and allocating the sensorson their xed positions in the structure to be monitored. The sec-ond procedure consists in the algorithm operation cycle.

In our algorithm we consider a hierarchical network topologycomprised of two layers. The lower layer contains sensor nodes, or-ganized in clusters, which are in charge of performing informationfusion techniques in data encompassed in the measurement and feature abstraction levels. The higher layer contains CHs in charge

of applying information fusion techniques in data at the feature anddecision abstraction levels. So, a cluster of sensors, composed by aCH and its subordinated sensor nodes, is considered the basicsensing unit in the network. Sensor nodes perform sensing tasksonly and CHs do not perform sensing tasks and are responsible forcoordinating the communication and processing inside theirrespective clusters.

5.1. Assumptions on the cluster formation

Before describing the Sensor-SHM procedures in details, it isimportant to mention that we assume that the WSN clusters arealready formed before the algorithm operation. The process of cluster formation and CH selection can be performed by a cluster-

ingprotocol,or manuallysetting (directlyover the images installedin the nodes). Sensor-SHM proper operation depends on the resultof the network clustering process, from which the algorithm usesthe outcomes to feed the arrays of sensors for each CH (detailedbelow). Sensor-SHM is agnostic to any specic clustering protocol.However, the criterion used to create clusters needs to be stronglyrelated to the SHM application needs. The number of clustersshould be preferably dened according to the number of structuralelements which we want to sense. The number of sensors per clus-ter should denote the amount of redundant data around one samestructural element (affects sensing accuracy and precision). CHsare vital for nding damage, so, the number of CHs is dened bythenumber of clusters (one CH per cluster) andtheir neighborhoodshould be set according to the structural properties of the sur-

rounding structural elements. For this reason, in spite of the factthat Sensor-SHM algorithmis agnostic to the underlying clustering



5/16

protocol, when choosing one of such protocol for using along withSensor-SHM, the best choice may be to adopt a semantic clusteringalgorithm [17], which considers the semantic relationships amongthe nodes positions and the sensing area. Choosing a clusteringprotocol like LEACH [18,19] is not a very suitable choice, sincethe random election of CHs may ignore the semantic correlationamong the sensor nodes and their positions in the monitored

structure. Also in LEACH, the number of sensors in each cluster ischosen to minimize the distance between sensors and CHs, anapproach that may assign sensors semantically correlated intodifferent clusters, causing the C j,t indexes to aggregate data fromdifferent semantic origins, i.e., different structural elements.Choosing a protocol as SKATER [20] is a more suitable choice,which still presents drawbacks. SKATER protocol assumes a net-work organization focused on data correlation among nodes.Therefore, the correlation among sensor nodes and their positionsis maintained indirectly, since the data collected by each sensor isdirectly correlated to the position of the sensor. However, SKATER algorithm still predicts a dynamic reelection of CHs, what mayimpose changes in the scenario used for analysis. In SHM applica-tions, any change in the spots chosen for sensing may provide anunreliable analysis, unless the semanticrelationships between sen-sors andtheir positions are strictly respected. The adoption of suchprotocols may impose a deep reformulation of the Sensor-SHMalgorithmfor dealing with such issues, but may also provide longernetwork lifetime. For this reason, theformulationof semantic clus-tering protocols may be an important contribution for the perfor-mance of Sensor-SHM algorithm. Nevertheless, the clustering andmulti-hop communication protocols may be freely chosen toimprove the communication capabilities of the sensors only. Thecoexistence of clusters for performing the application require-ments (Sensor-SHM clusters) and clusters for performing reliableand efcient data transportation only may result in a general per-formance increase, at the expense of energy.

5.2. Setup procedure

In Sensor-SHMthe structural monitoring is performed in a peri-odic basis, and each monitoring cycle is based on a collected signa-ture sample. Since the data collection ( measurement level ) is theoperation which takes the longest time to complete during theoperation cycle of the algorithm, the operation cycle of the algo-rithm is referred by its respective data collection stage . A data col-lection stage is identied by an integer t , which is incremented byoneforeach performed data collection stage. Thesestagesstart at agiven time, dened by the sink node. As an example, in our exper-iments the collection period , which represents the duration of each data collection stage, is dened as being long enough to col-lect 512 acceleration samples at a sampling rate of 1.0 kHz, result-

ing on a collection period which lasts for about 500ms. Thesenumbers are not xed in the algorithm. They can be dened foreach application. In general, the number of collected samples isdetermined by the following criteria: (i) it must be enough to en-sure a good resolution in thepower spectrumthat will be returned,which implies in better precision in the modal frequencies deter-mination, (ii) it must be a power of 2, since this is a requirementfor the entry of data in the FFT algorithm, and (iii) it shall notexceed the sensors storage capacity (Flash memory). The sampling rate is set according to the following criteria: (i) it must be greaterthan the value of the rst modal frequencies of interest so thatthese are shown in the power spectrum (commonly, values below200 Hz for the rst ve modal frequencies of structures are ex-pected), (ii) it must be high enough to ensure accuracy, (iii) it

should be twice the highest modal frequency of interest, to meetthe Nyquist criterion. In [3] a sampling rate of 1.0 kHz is used, with

hardware similar to the one used in our work, showing that it ispossible to achieve this sampling rate in a practical situation.

To ensure thesynchronization among thedata collected by eachsensor, each data collection stage should start at the same time onallsensors, so that there is meaning in thecomparisonof signaturesof this stage and earlier stages, and in comparison of signaturesfrom different sensors. The synchronization problem is reported

and addressed in [3,9]. A feasible network implementation to dealwith thesynchronization issue is to let thesink nodebe responsibleforsending a message to thewholenetworkscheduling thestartingtime of the next data collection stage, assuming that all sensorshave their internal clocks synchronized.

Other parameters, such as the array of CH neighbors and the ar-ray of sensors that are part of a cluster, must be set. The array of CHneighbors informs each CH who are its neighbors. Neighbors of aCH are CHs, from other clusters, which are allowed to communi-cate among each other in order to accomplish the tasks relatedto damage localization and extent determination. For each CH,the array of sensors informs which are the sensors subordinatedto it. When using both arrays, all the necessary communicationscan be easily established. The sets of existing sensors and CHsare properly dened as a collection J of Z CHs where each CH isidentied by j = {1, 2, , Z}, and has a subset of subordinated sen-sors. All the subsets of sensors subordinated to each CH are part of another collection dened by I , that includes all the N sensors inthe network, where each sensor is identied by i = {1, 2, , N}.Constants like T i, L j and Ai must also be set before the deployment,although they can be changed during the network operation. Thedenitions of these constants andtheir useareexplained in thefol-lowing subsections. These constants are stored in the CHs.

After all these settings, the network can be physically deployedover the sensing area (the chosen structure), and the nodes can beinstalled in their xed positions. As part of the initialization pro-cess of the sensor nodes, they collect the initial signature of thestructure, x i,0. It is supposed that at this time the structure is atthe beginning of its operation. Thus, each sensor generates itsx i,0 vector and transmits it to its CH. These initial values will beused as a reference for the undamaged (healthy) structure. Then,sensors enter in sleep mode and wait for the next data collectionstage, which will be identied by a value of t = 1.

5.3. Data collection stages from the sensors viewpoint: measurement and Feature levels

The data collection stage starts at a given time, as requested bythe sinknode. A message is sent fromthe sink to the CHs, and thoseare responsible for sending messages to schedule the next sensingtask on their subordinated sensors. Then, when the specied timecomes, the sensors start collecting data. So, at every data collectionstage t , each sensor node collects the acceleration data in the time

domain at its relative position andperforms a FFTon therespectivecollected signals, our rst DAIDAO information fusion. Then asimple method is applied to extract the modal frequencies in thepower spectrum generated in the previous step, what can be con-sidered a DAIFEO information fusion. This method has the samegoal of the Polynomial Curve Fitting seen in [14], but is much lessexpensive in terms of energy, and is able to be fully implementedwithin the sensors. However, it is less accurate, leading to moreerrors. This lack of accuracy can be balanced by increasing thenumber of sensors in each cluster, generating more redundantdata.

The frequency values extracted from the power spectrum gen-erated by the FFT, assuming no noise interference, are related tothe modal frequencies of the structure, and are the rst useful fea-

tures over which the CHs can take decisions. Formalizing, x i,t isasignature of a given structure, acquired at a data collection stage



6/16

t , and represented by a vector of M sensedfrequencies in a sensoridentied by i. Therefore, different sensors get different values of signatures for the structure, depending on their location and theinstant of time in which the data collection stage started. For therst 5 modal frequencies ( M = 5), the vector has the form asdescribed in Eq. (1). Finally, all vectors generated by each sensorat each data collection stage t are sent to their respective CHs. This

nal step at the feature data abstraction level signals the end of thenode activity for this data collection stage, and it goes into sleepmode.

5.4. Data collection stages from the CHs viewpoint: feature anddecision levels

The CHs are responsible for performing the following steps:damage detection and determination of damage location andextent.

5.4.1. Damage detectionThe CH is responsible for comparing the x i,0 vectors from all its

subordinated nodes and the subsequent x i,t vectors, generated in a

similar way as x i,0 in the further data collection stages. The veri-cation of change in the modal frequencies of a structure is per-formed by comparing x i,0 and x i,t vectors. The comparison isdone using the absolute value of the difference between x i,0 andx i,t , and the result is stored in the D x i,t vector, as seen in

x i; t

x 1i; t

x 2i; t

x 3i; t

x 4i; t

x 5i; t

266666664

3777777751

D x i; t jx i; 0 x i; t j

jx 1i; 0 x1i; t j

jx 2i; 0 x 2i; t jjx 3i; 0 x

3i; t j

jx 4i; 0 x4i; t j

jx 5i; 0 x5i; t j

266666664377777775

D x 1i; t

D x 2i; t D x 3

i; t

D x 4i; t

D x 5i; t

266666664377777775

2

For D x i,t values different from the null vector, the CH can as-sume, considering a given T i tolerance value, that there has beena signicant change in the structure, which may indicate the pres-ence of damage or the action of a temporary external event. If avalue from one of the D x i,t vector positions exceed its tolerancefor a sensor i, the CH proceeds to the next step in the monitoringprocess, which refers to the damage localization, since it has de-tected an abnormal condition in the structure. This decision is

theoutput of a FEIDEO information fusion andreduces the energyconsumption since it avoids thewasteof energyfrom the followingunperformed steps. The tolerances for the D x i,t vector arestored inthe T i vector. It is important to mention that the T i vector is deter-mined for each sensor based on knowledge and analysis of thelocalities in which each sensor will be installed. Also, the T i vectorcan be statistically determined after making a series of experimen-tal samples. The purpose of adopting a T i tolerancevector is to pre-vent small random disturbances, which do not imply theoccurrence of abnormal conditions, from being considered by themonitoring procedure as such.

5.4.2. Damage localization and extent determinationKnowing that changes in the frequencies of the higher vibration

modes mean changes in local vibration modes, each CH analyzesthe D x i,t vectors of all sensors located in its cluster in search for

these kinds of changes. In each CH and for each data collectionstage t , this analysis is performed with the help of the Di,t coef-cient, which is calculated for each sensor i that has exceeded thegiven tolerance, in the given cluster (Eq. (3)). The Di,t coefcientis set so that its value indicates how close the sensor i is fromthe damaged site, and Ai is a vector of weights, assigned to eachmodal frequency. To identify the sensors that are closest to the

damage site, highervalues to theweights associated with the high-er modal frequencies can be assigned.

Di; t AT i

D x i; t A1

i A2

i A3

i A4

i A5

i

D x 1i; t

D x 2i; t

D x 3i; t

D x 4i; t

D x 5i; t

266666664

3777777753

C j; t Xk

i1Di; t 4

So, according to Eqs. (3) and (4), Di,t and C j,t coefcients are out-puts of the FEIFEO information fusion. To sum up, the weightsshould be assigned so that sensors belonging to a given clusterlocated near the damage site obtain the highest Di,t coefcientsof the whole network. In other clusters, the Di,t coefcients shouldbe smaller, but still nonzero, since the lower frequency values willbe changed and these changes are identied by many other sen-sors. In the following step, Di,t coefcients are aggregated in eachcluster j, by summing their values for all k sensors in the cluster,resulting in a C j,t coefcient (Eq. (4)). By its mathematical deni-tion, C j,t coefcient is an indicator of how close to the damagethe cluster as a whole is. The algorithm uses this indicator to locateand determine the damage extent.

Our algorithm of damage localization and extent determinationis classied as a FEIDEO information fusion. In this algorithm,each CH node compares its C j,t coefcient with a L j tolerance value.When the C j,t coefcient exceeds L j, the CH sends a messageinforming its C j,t coefcient to its neighbor CHs. The L j toleranceis dened for each CH, in a similar way to the determination of the values of the T i tolerance vector. The tolerance values dependon the structural characteristics, and therefore should be deter-mined by an expert in the structure, and through statisticalanalysis.

After the CH j transmits its C j,t value to its immediate CH neigh-bors, it is expected that some of these neighbors also haveexceeded their L j tolerance value, and thus have sent their respec-tive C j,t coefcients to their neighbors. CHs then compare receivedC j,t values with their own and the CH who has the greatest C j,t value

in a given neighborhood assumes the role of a collector. The col-lector node is responsible for two tasks: (i) aggregate the informa-tion about the x i,t values from all its neighboring CHs and build areport to be sent to the sink, issuing a warning and (ii) act on theenvironment around it by triggering a relay, aiming to preventthe progression damage and avoiding further problems in thelocality. The collaboration among CHs can also be classied as aDEI-DEO information fusion, since each CH has its own local deci-sion for assuming damage or not, but it is allowed to consider thedecisions of other local neighbors to present a more reliable deci-sion. To build the report that will be sent to the sink, the valuesof x i,t vectors at that data collection stage were chosen, becausefrom these values it is possible to deduce all the other relevantinformation. Since the sink node has knowledge of x i,0 vectors,

and all the weight and tolerance values, it is possible to calculateall the other related values that were previously shown in this



7/16

explanation, and still have a global view of the events that hap-pened into the network during the data collection stage t .

It is assumed that the damage location and extent are deter-mined by the positions of the sensors that are CHs and whose C j,t coefcient exceeded the L j tolerance value at a certain moment.In case of multiple damage sites, or large damage sites that covera large area on the structure, the trend is that there will be many

emerging collectors, and multiple reports from different locationswill arrive at the sink node. It is important to note that the sinknode only plays the role of supporting the CHs in their tasks, send-ing a message to the whole network, scheduling the time of datacollection stage, storing historical data over time and makingrequests that represent interventions on the network. No calcula-tions relative to damage detection, localization and extent deter-mination must be made at the sink. At the sink, the damagelocalization is done through the unique identiers and positionsof the reporting sensors, and the extent is determined by the areacovered by these sensors. Through the values contained in thereport, it is possible to reproduce the situation that occurred with-in the network and take the appropriate decisions about the struc-tural predictive maintenance.

6. Experiments with Sensor-SHM

This section describes the experiments conducted to evaluatethe performance of Sensor-SHM in two main ways: (i) the algo-rithm capability to precisely and accurately detect, localize anddetermine the extent of damage sites, and (ii) the algorithm ef-ciency when running within a resource constrained WSN, in termsof communication. Regarding item (i), the efciency of the algo-rithm will be evaluated by using sensor nodes that sense simu-lated acceleration data instead of using real, sensor-collected,acceleration data. Regarding item (ii), in our previous work [7]we presented an energy consumption analysis and a partial com-munication analysis, in which we achieved promising results. So

in this paper our focus regarding item (ii) will be on providing amore comprehensive analysis on the impact of the algorithm overthe communication. Two experiments were set up to evaluate item(ii), (Sections 6.2, 6.3 and 6.4) and one experiment was set up toevaluate item (i) (Sections 6.5, 6.6 and 6.7), totalizing threeexperiments.

6.1. Network prototyping

In all the performed experiments we prototyped a sensor net-work composed of MICAz motes from Crossbow Technology [21].The motes are programmed in nesC language, under the TinyOS2.1 development environment [23]. The implementation of ourSensor-SHM algorithm encompasses two programs; one for run-

ning inside motes assigned as CHs and other for running in motesacting as sensor nodes. The default implementations of the802.15.4 protocol for lower level communication handling, andthe Active Message protocol [22] for higher level communicationhandling in TinyOS 2.1were used. Our experiments aimat evaluat-ing the performance of the Sensor-SHM algorithm alone, so, a leanimplementation of the whole system in our prototype was desired.

From the point of view of the sensor nodes, the algorithm isstructured in four main components ( Fig. 2): SensorMainC , Sensor-CollectC , SensorFFTC and SensorRadioC . Besides them, basic Tiny-OS components were used to implement radio and sensor boards[24] (they are hidden in the gure for the sake of clarity). The MainC component provides the Boot interface, and it is theprimary component of TinyOS, from which the nodes boot (initial-

ize). SensorMainC component manages all the sensor nodes in thenetwork. To achieve its goal, SensorMainC makesuseof SensorCol-

lectC and SensorRadioC components accessedrespectively throughthe SensorCollect and SensorRadio interfaces. SensorCollectC con-trols sensing tasks, using other basic interfaces provided by TinyOS2.1. SensorCollectC also makes use of the SensorFFTC component,through the SensorFFT interface, which is responsible for perform-ing a FFT within the wireless mote. SensorFFTC was built as a sep-arate component to enable thereuse in other applications, given its

generality and likely use in different domains. SensorRadioC isresponsible for the communication between each sensor and itsCH, and uses the basic TinyOS 2.1 radio interfaces, based on ActiveMessage [22].

From the point of view of the CH nodes, the Sensor-SHM algo-rithmencompasses threemain components ( Fig. 2(b)): LeaderMa-inC , LeaderApplicationC and LeaderRadioC . LeaderMainC component manages the Sensor-SHM algorithm in the CH nodes. LeaderApplicationC is used to calculate the C j,t coefcients forthe cluster, based on the values of natural frequencies receivedby the radio. The CH radio is managed by the LeaderRadioC , whichuses the basic TinyOS 2.1 radio interfaces as the SensorRadioC component. However, the LeaderRadioC has a larger and more de-tailed implementation than the SensorRadioC , since the CH makesa more intensive use of its radio. For instance, the LeaderRadioC isin charge of requesting and collecting the frequency values of allsensors within its respective cluster, while the SensorRadioC com-ponent implements only the response to the received requestsfrom the single respective CH.

This separation is important since our network topology is sta-tic and, therefore, the CHs are always the same nodes (we do notassume the CH rotation). In this situation a CH node will never as-sume the roles of a sensor, and vice versa. So, to avoid waste of memory in the program used by the sensor nodes, only the tasksconcerning the sensor node side in communications are imple-mented for their radio component, which are simpler than thetasks regarding the CH node side in communications.

6.2. Methodology to evaluate resources consumption

Two experiments were performed in real sensor hardware, aim-ing to evaluate Sensor-SHM in terms of communication, meaningspecically communication overhead and packet loss . To evaluatethe communication overhead, we measured the number of pack-ets, and the number of bytes, both transmitted by sensors andCHs, analyzing separately data messages and control messagestransmissions. We aim at understating the relation that governsthe communication overhead by increasing the number of clustersand sensors.

Data messages are messages whose payload contains informa-tion on the structural state. There are two kinds of data message inour implementation: (a) messages to convey information on thevalues of natural frequencies measured by the sensors (mainly

exchanged between sensors and their CHs), and (b) messages tocarry information on the values of C j,t computed by CHs (mainlyexchanged among CHs). Each data message of kind (a) requiresfour bytes per frequency in its payload. Once we use 5 natural fre-quencies, it means that this kind of data message has a 20 bytespayload. Data messages of kind (b) require 4 bytes in its payloadto transport a value of C j,t .

Control messages are messages with empty payloads, for gen-eral (oftenmanagement) purposes. Their main uses are to establishcommunication among the nodes and to disseminate commandsinto the network, for instance, to start the data collection stages.Control messages are exchanged among all nodes of the network.

We considered an indoor environment, using MICAz motesincluding only the mote board containing the processor, radio,

memory and batteries. These items are depicted in [21]. We didnot use sensor boards, once the sampled data were simulated in



8/16

this set of experiments. The acceleration values collected by thesensors were the same at every data collection stage and the rulesto set the collected values were as presented in [7]. So, since we arefocusing on the communication overhead generated by the rawdata sampling, the meaning/semantics of acceleration valueswhich are being collected is irrelevant. Five summed sinusoidswith known frequencies made every sensor return modal frequen-cies of 20 Hz, 40 Hz, 60 Hz, 80 Hz and 100 Hz. The range of 100 Hzis where themost important natural frequencies of large structureslay in, thus it is possible to evaluate the precision of the FFT andpeak extraction algorithms over this range.

During the data collection stages, each sensor generates a stan-dard error (standard deviation) of 2 Hz in the determination of the modal frequencies due to imprecisions in calculations (mainlyrelative to the own FFT algorithm and the peak extraction algo-rithm imprecision) and truncations (which are needed to transmitthe data over the radio). All the tolerances mentioned in the algo-rithm description (see Section 5) were set to zero, so we assumedthat in all thedata collection stagesdamage sites were found, whatis considered to be the most resource demanding scenario. More-over, for the whole set of experiments, we used the default valuesfor the MICAz radio parameters [21]. Therefore, the maximummessage payload size was set to 28 bytes, and the radio data ratewas 250 kbps. No sleep mode was implemented, so the radio dutycycle was 100% since the experiments were short and the amountof energy spent during the idle time is very low, thus not interfer-ing with the results.

6.3. Description of the scenarios used to evaluate resourcesconsumption

The pair of experiments for evaluating communication is re-ferred as experiments (I) and (II). In experiment (I), each scenariois characterized by the variation in the number of nodes in thesame cluster, to evaluate communication inside one cluster(intra-cluster). In experiment (II), each scenario is characterizedby the variation in the number of clusters, keeping a xed numberof sensors per cluster, to evaluate communication among CHs(inter-clusters). Both experiments had 4 scenarios, which aredescribed basedon a two-dimensional Cartesian plane ( x, y), wherethe sink node is always at the origin.

In experiment (I), all the scenarios had onlyone CH. This CHwasalways placed at 1 m fromthe sink node in the x axis. The rst sce-nario had one sensor pertaining to this cluster, at 1 m from the CHin the x axis. The next three scenarios had the number of sensors

pertaining to this cluster increasedby one, following a linear topol-ogy along the x axis, spacing each sensor 1 m from the last one

added. In experiment (II), the rst scenario is the same as in exper-iment I (one cluster, containing one CH and one sensor). The num-ber of clusters is increased by one in the next three scenarios,equally spacing the CHs 1 m from the last one added in a lineartopology in the x axis. The single sensor is placed at the same xposition of its respective CH, and 1 m from its respective CH inthe y axis. In all scenarios of both experiments, all sensors werewithin radio range of the sink node, considering the 20 m to30 m indoor range of the MICAz motes in [21]. In every scenarioof experiment (ii), all the CHs were considered neighbors amongthemselves.

In our previous work [7], we found that each data collectionstage takes around 10 s to complete, and the number of sensorsin the network did not have a signicant inuence over this time.Then, for the experiments in this work, the chosen time betweeneach data collection stage was 15 s, to assure that all data collec-tion stages are nished and, at the same time, to minimize the idletime between each data collection stage. So, at every 15 s the sinksends a message for each CH to start a data collection stage. Tendata collection stages were performed.

6.4. Results and analysis of resources consumption

The rst observation from the results of the set of experimentsdescribed in Section 6.2 (Fig. 3) is that the number of transmittedbytes related to Data Messages is bigger than the number of trans-mitted bytes related to Control Messages in both experiments. Thisresult points towards a successful lean implementation of our pro-totype, due to the achieved low communication overhead causedby control mechanisms or auxiliary protocols, as we intended.We also consider our implementation successful, due to the low

packet loss rate achieved among the sensor nodes in the network,what contributed for the quality of the analysis of the structuralintegrity. In fact, no retransmissions were needed among nodesin both experiments. This fact is explained by the proximity amongnodes considered in each topology, assuring a good link qualityduring transmissions. Thisproximity reects a harshscenario sincein SHM applications the usual distance between motes is evensmaller than 1 m, as it can be seen in [14]. But we have faced apacket loss rate of 3.6% between the sink node (which was promis-cuously listening to all the network trafc) and the rest of the net-work. We attributed this outcome to a lack of computationalpower of the sensor node at the sink, which could not deal withthe transmission of packets over the serial port at the same ratewith which it received them from the radio. At certain moments,

many nodes transmitted packets at the same time, and the buffersize at the node was not enough to deal with all messages. This

Fig. 2. Component diagrams for (a) sensor and (b) cluster-head nodes.



9/16

packet loss rate did not represent a signicant interference in ourresults, since we could t curves to our experimental data afterthe experiment and easily detect a lost packet in our controlledenvironment. But this computational limitation can be resolvedthrough the adoption of a 2-tiered topologyin which thesink node,or even the CH nodes, could detain a greater computational power.

In experiment (I), for every new sensor added to the cluster, thetotal number of transmitted bytes increased linearly. This totalnumber is the sum of transmitted bytes due to both Control andData Messages. The CH was the only node which had its numberof transmissions changed due to the increasing of the number of sensors and this change was in the number of transmitted ControlMessages. The number of Control and Data Messages transmittedby each sensor and the number of Data Messages transmitted bythe CH were not affected by the change in the number of sensorsin the same cluster.

In experiment (II) for every newcluster added, the total numberof transmitted bytes increasedfollowing a polynomial relation.Theshape of this curve is mainly explainedby the number of transmis-sions of Data Messages. And the variation in the number of trans-missions of Data Messages is explained by the number of DataMessages exchanged among CHs. Since every new CH added wasconsidered neighbor of all theCHs already existent in theprior sce-narios, the number of Data Messages (used to transmit the valuesof C j,t among CHs) increased at a much faster rate. The number of

Control and Data Messages transmitted by each sensor and thenumber of Control Messages transmitted by the CH were not af-fected by the change in the number of clusters.

The formulas in Fig. 3 are good predictors for the number of transmitted bytes in a situation comprisingmore nodes. Also, sincethe number of transmitted bytes for each CH changes (increases)with the addition of more sensors per cluster or more neighborclusters, the CH may be overloaded at some number of sensors inits cluster, or neighbor clusters. We did not have the sufcientnumber of sensor nodes to reach this limit experimentally. Bythe obtained results, our prototype presented a good scalabilitywhen considering the expected number of sensors to be used inmost currently available related applications in a real situation.This maximum number of sensors in SHM applications found byus in the literature was 64 [3]. The execution time of the data col-lection stages starts to increase when the number of bytes neededto be transmitted by the CH reaches the radio data rate (250kbpsforMICAz radio). And increasing the number of sensors in the samecluster is a faster way of reaching this limit than increasing thenumber of neighbor clusters.

It is also important to mention that the number of bytes trans-mitted by a sensor node is a good estimation for the energy spentby it. Since it is well-known that radio transmissions are the mostenergy consumptive tasks in WSNs [14], it is possible to estimatethat CHs will have a shorter lifetime when more sensors are added

Fig. 3. Results for experiments (I) and (II).



10/16

to its cluster. So, in our prototype, it is always a better choice toallocate the number of needed sensors into as many clusters aspossible. This result is also a strong justication for the use of CHs andthechoice for distributing the algorithmamong the nodes,making it localized. But still, this certainly increases the number of nodes needed for the same task, what may increase theoverall costof the system. So, there is a tradeoff to be managed and experi-

ments and simulations performed previously to the WSN deploy-ment are very useful to guide the design decisions.

6.5. Methodology for accuracy evaluation over simulated raw data

In this experiment focus was given in the ability of the algo-rithm to correctly identify a damage site and its extent. The inputdata about location and extent of damage was introduced throughchanges in the simulated raw acceleration data collected by thesensors. Although this experiment is performed over simulatedraw acceleration data, it is fully able to show how Sensor-SHM isexpected to behave in a real-world situation, and is also a rst testfor assessing theprecision of themathematical model to be used infuture works. The hardware used in this experiment was the sameused in the previous one, based on the MICAz motes. The networktopology is shown in Fig. 4.

Four clusters are considered in this topology, which are com-posed of two sensors and a CH each. Also, every cluster is consid-ered to be deployed over a different homogenous structuralelement. Each of these elements could be represented by a xedbeam, and all the four beams are connected among themselvesby structural elements which will retransmit vibrations. The sen-sors are considered to be directly attached over the structure, butthe CHs do not need to be attached. Each cluster-head must onlybe within radio range of all its neighbor (single-hop) cluster-headnodes. Sensors that are over the same homogenous structural ele-ment are expected to produce more redundant data amongthemselves.

Based on the experience acquired during our experiments, wehave found different methods to represent an unhealthy state of the structure in this simulated scenario. The most relevant of themare (i) increasing the value of intensity for one of the modal fre-quencies, (ii) adding a sixth modal frequency ( x 6) among the exis-tent frequencies, or removing one of the veexistent ones, and(iii)shifting the values of one or more frequencies. Methods (i) and (ii)wouldnotbe directly perceived by Sensor-SHM, but their collateraleffects can still be perceived. In (i) one of the peaks of the powerspectrum can mask the other peaks, e.g. if it is the third peak, itmay be perceived as the rst and probably the only one, what willbe interpreted as if the rst peak hadbeen shifted from its value to

the value where previously was the third peak. In (ii), adding orremoving one of the peaks among the rst ve would also makethe peaks to be perceived in a wrong order, generating a similar ef-fect than in (i). So, the method (iii) is chosen since it directly causesthe expected changes and is easier to debug in Sensor-SHM.

6.6. Description of parameters used to simulate raw data

The simulated raw acceleration data correspond to an acceler-ometer output in the z axis, named a z ( x). The general rule to gen-erate a z ( x) is dened by Eq. (5) as a damped linear combinationof ve sinusoids weighted by the respective intensity ( I m), i.e. theamplitude of each modal frequency in the signal.

a z x e0

:5 x X

5

m1I m sin 2 p x m x" # 5

In Eq.(5), the time ( x) variesaccordingto the time period relatedto the sampling rate, which was chosen in this experiment as1.0kHz. The initial value for each parameter in each node is:x 1i; t 20: 0 Hz; x 2i; t 40: 0 Hz; x 3i; t 60: 0 Hz; x 4i; t 80: 0 Hz; x 5i; t 100: 0 Hz; and the intensities were all set to 0.25, regardless of themode of vibration. The values of intensity are all the same, sincewedo not want tomaskfrequenciesin thespectrum, withsome fre-quenciesbeingmore dominant than others. Also, the intensities re-ceive small values, generating datain the range thataccelerometersareable to collect in real situations.Through numericalsimulationswereached to thevalue of 0.25. This value generatesa signalwhichcan be reasonably dealt with in our peak extraction algorithm. Theresults of the experiments from [14] present acceleration data sim-ilar to the one which we generated by Equation (5). The initial val-ues of frequencies were the same as described in Section 6.2.

To generate the frequency shifts, new ( Bmt) factors were in-serted in the equation of each sensor node as shown in Eq. (6).

A z x e 0 : 5 x

Xs

m1 I m sin

2 p

x m

Bm t x" # 6

These factors are all functions of two different parameters: (i)thenumber of thecurrent data collection stage t , and (ii) a constantof magnitude Bm. In Eq. (6), to make each frequency to shift at dif-ferent rates, we added the variable t along with the Bm magnitudevalues. So, at every data collection stage, the variable t increases byone, and for each node, each frequency will be increased by thesmall value of its respective ( Bmt) factor which increases at eachdata collection stage as a multiple of Bm . The values of Bm for eachnode are created according to the following criteria: (i) it isexpected that Bm is higher for frequencies of the vibration modeswhich are more sensitive to damage at the node locality, i.e. thehigher modes and (ii) the values of Bm are normalized, so the high-

est Bm value is 1, for the most sensitive mode of the node which isclosest to damage. To perform a random damage generation, fre-quency variation patterns in presence of damage were dened asshown in Table 1.

To set up the values in Table 1, we rstly assigned the highestvalue of Bm to the highest mode of vibration of nodes in the clusterwhich is closest to damage, so the rst chosenvalue is B 5 to patternA. Next, we assigned decreasing values to the other magnitudes of

Fig. 4. Topology used for accuracy evaluation over simulatedraw data. Each node isrepresented by a circle, and uniquely identied by an ID.

Table 1

Frequency variation patterns in presence of damage.

Variation pattern B 1 B2 B3 B4 B5A 0.2 0.4 0.6 0.8 1.0B 0.2 0.2 0.3 0.4 0.5C 0.2 0.0 0.0 0.0 0.0



11/16

pattern A, using an arithmeticalprogression of ratio 0.2. Next, sincethe B1 mode is considered global, and is perceived in the same wayby all sensors, the column of B 1 received the same value for alldamage patterns. The last values of pattern B are the values of pat-tern A divided by 2, since pattern B will be assigned to nodes thatare in the neighborhood of the damage site, whichwill receive pat-tern A. Except for B1 mode, pattern C receives zero values for the

other modes, since this pattern will be assigned to nodes whichare far from the damage site, and will perceive shifts only in theirglobal frequencies. Table 2 presents all the possible cases of dam-age pattern assignments for our experiment. Once congured atthe network setup, the assignments remain the same until theend of the experiment. This assignment is randomly made andeach damage pattern has 25% chance of being chosen.

Since at rst sight it is not possible to know where the damagesite will be, the rules to choose the values of Ai are based in theknowledge of the general rule which states that higher frequenciesare more sensitive to damage. In order to speed up the processes of damage detection, localization and extent determination, highervalues to the higher frequencies are assigned, so C j,t will dependmore on the variations of the higher frequencies. Also, the valuesof Ai are normalized, in which the higher value is assigned to thefth mode, and decreasing values are assigned to the other modes.The rule is similar to the one used to create pattern A in Table1. So,for every sensor i, Ai = {0.2;0.4;0.6;0.8;1.0}.

After performing preliminary tests and numerical simulations,by adopting an empirical methodology (based on trial and error),we found out that setting the value L j = 20.0 for all clusters is en-ough to prevent a high occurrence of false positives, especially inthe rst data collection stage, and to assure that damage will onlybe signalized when the frequencies have expressively shifted, incomparison to their initial values. So, for every cluster j, L j = 20.0.The values of T i are the same for every node and every modal fre-quency. The reasonfor this assignment is to prevent false-positives(see Section 5.4.1) since each sensor generates a standard errorof 2 Hz in the determination of the modal frequencies due toimprecision in calculations and truncations. So, T i will assumethe value of two standarddeviations. This is enoughto assume thatin a real situation less than 5% of frequency values will exceed thelimit in an undamaged situation (false-positive), assuming that theextracted values of frequencies follow a normal distribution andthe average is well estimated using 7 frequency samples. Then,the T i limit assumes the value of 4 Hz for every mode of vibrationand every node.

Finally, the values of Tables 1 and 2 are stored within the sen-sors, and used to generate the random data. Our goal is to observethe behavior of C j,t and the x i,t metrics during the data collectionstages, i.e. a sensitivity analysis of the indicated variables will beperformed over the variation of the acceleration signals collected.

6.7. Results and analysis of accuracy over simulated raw data

The network performed the setup and 29 data collection stagesin this experiment stopping after enough time for damage to beclearly characterized in the scenario. Our program that runs onthe sink was set up to randomly choose one of the four cases in

Table 2. Case 1 was randomly chosen to be performed for thisexperiment. A packet loss rate of 0.4% for the data messages con-taining frequency values was detected. The explanation for thisfact is the same from the previous experiments (Section 6.4). Forthis experiment, the values of frequencies missing due to this rea-son were interpolated after the end of the experiment.

The peak extraction algorithm performed poorly. During the

whole experiment, around 15.2% of the frequency values couldnot be correctly found. When the peak extraction algorithm doesnot nd the current frequency value correctly, it assumes that itsvalue is equal to the respective initial value stored in the x i,0 vec-tor. For instance, considering all the extracted frequency valuesfrom the respective mode in the experiment, the rst mode wasnot found in 12.5% of the cases, the second 5.0%, the third 20.8%,the fourth 7.5%, and the fth 30.0%. It means that the fth peakwas the most difcult to nd and the best performance wasachieved over the second peak. Such results pointed out that ourpeak extraction algorithm still needs improvements to performwell in terms of accuracy. So, we conclude that it may not be agood choice to generalize the use of this algorithm for every struc-ture through the presented methodology of peak extraction.Instead, it may be a better choice to make prior studies about spe-cic modes of vibration from the structure to be monitored, so thatit is possible to indicate more parameters to the peak extractionalgorithm, what will lead to an improvement in its accuracy. Thisfact does not invalidate the algorithm; it only imposes a restrictionto the methodology of data acquisition.

We chose a node which received the variation pattern A (NodeID 1), one which received variation pattern B (Node ID 3), and onewhich receivedvariation pattern C (Node ID 5) to showin Fig. 5 theevolution of each frequency shift for each variation pattern.Wecansee the frequencies following linear trends. The angular coefcientis exactly the value of Bm. In fact, this mathematical relation helpsunderstanding our proposal of damage simulation presented inSection 6.6. When changing the term x m from Eq. (5) tox m Bm t in Eq. (6), we imposed to the frequencies a lineardependency over the data collection stages. In Node ID 5, sinceits values for Bm are 0.0 for the last four natural frequencies, theselast four natural frequencies followed a constant trend and werealwaysoscillating into their control limits of T i = 4 Hz. Itis also pos-sible to observe that the three nodes in Fig. 5, and all the othernodes which received any variation pattern, perceived damage inthe lowest frequency, at the same rate. As expected, nodes whichreceived variation pattern A had greater variations in their fre-quencies than the other nodes during the experiment. It is alsoimportant to understand the meaning of the peaks pointing downwhich are shown in almost all curves from all graphs. They repre-sent moments in whichone of the frequencyvalues is not correctlyextracted by the peak extraction algorithm. As previously men-tioned, when this case happens, the algorithm assumes the value

of x i,0 for that natural frequency. In the following data collectionstages, it returned to the expected linear trend. When the valuewas not correctly extracted, it had implications in the calculationsof C j,t , and Fig. 6 illustrates this issue.

Fig. 6presents a graph with the evolutions of C j,t values with theL j tolerances evidenced during the data collection stages for thefour clusters in the experiment. Five important situations are sig-nalized in Fig. 6, whichneed to be discussed. Situation1 points thatthe rst cluster to detect damage was cluster 1, at data collectionstage 4. At this time, this cluster could be signalizing a false posi-tive, but at stage 5, this same cluster detects damage again. So, itis possible to assure that damage really occurred close to cluster1. Situation 2 points that damage had progressed so much thatcluster 2 started detecting it, in smaller proportions. It is the con-

cept of damage extentdetermination. Two clusters arenow detect-ing damage, the one closer to damage with higher C j,t values.

Table 2

Possible cases for frequency pattern assignments.

Node ID Case 1 Case 2 Case 3 Case 4

1 and 2 A B C C3 and 4 B A B C5 and 6 C B A B7 and 8 C C B A



12/16

Situation 3 concerns to the problem detected in Fig. 5. When thefrequencieswere not correctlyextracted, thevaluesof C j,t were dis-turbed, but since C j,t depends on shifts of the ve natural frequen-cies, losing only one value of natural frequency is not enough tobring C j,t below the L j line. We therefore observed that to increasethe reliability of our algorithm it is better to consider in the anal-ysis as many natural frequencies as possible, so that if one of thesefrequencies is lost, or wrongly estimated, the disturbance over C j,t will not be enough to harm the accuracy of the algorithm.

Situation 4 shows the rst time in which one of the nodes fromthe clusters with similar behaviors (3) and (4) present a value forone of its natural frequencies that surpasses the tolerance T i. Thisnatural frequency is the rst one which started at 20 Hz. This dis-turbancein therst natural frequencywas enough to cause thecal-culation of C 3,t for the rst time in moment t = 18 and C 4,t inmoment t = 21. At a future data collection stage, if the input param-eters remained, clusters 3 and 4 would start signalizing damage,only due to therst natural frequency shift, another case in damage

extent. The network did not go that far in timeduring its operation,since it would take much time to observe this fact.

Finally, situation 5 shows the effect of three consecutive stagesof data collection in which the fth natural frequency was wronglyextracted over C 1,t . The fth natural frequencies of nodes 1 and 2are the ones which are most sensitive to damage, so, not extractingthem correctly made the values of C 1,t drop below 50% of itsexpected value. If L1 were set higher than 80.0, cluster 1 wouldnot signalize damage detection, when actually the damage is stillprogressing. If this state takes too long, this would be the case of a false negative. The key to avoid this issue is to rely on the opin-ions of the neighbor clusters. In this case, cluster 2 would keep sig-nalizing damage if its limit L2 were set to 20.0. Therefore, falsenegative avoidance is one of the main reasons for stimulating col-laboration mechanismsamong thenodes in this algorithm. The im-pact of packet loss between a sensor and a CH for the proposedmethod is similar to the impact of a wrong peak extraction, aspointed in situation5. But when a packet is lost, all the ve natural

Fig. 5. Variation of frequencies for Node IDs 1, 3 and 5, during data collection stages.

Fig. 6. Variation of C j,t values per cluster during data collection stages.



13/16

frequencies are lost, and therespective C j,t coefcient tends to droplower than in the case where only one frequency is wrongly ex-tracted from the power spectrum. In case of a packet lost in thecommunication among CHs, the whole region monitored by therespective cluster is instantly out of the analysis, what may deeplycompromise the whole analysis. For this reason, the adoption of protocols to provide as much reliability as possible (for instance,

by using acknowledgments of messages) in the communicationamong CHs is desired.

6.8. A case study for assessing the accuracy of the algorithm

In this section we present and detail a procedure to deploy awirelesssensor networkcomprised ofnodes runningouralgorithm.Theprocedure is basedon the real experiment described in [8] (seeSection 4). The scenario parameters were congured so that thebehavior of our algorithm could be analytically simulated and ana-lyzed. This case study presents a possible real application for Sen-sor-SHM algorithm, and allows analyzing its performance interms of accuracy on assessing the integrity of a structure. We ex-tracted the data about the natural frequency changes due to dam-age occurrence from the test performed in [8], and prepared anumerical simulation for assessing the behavior of Sensor-SHMalgorithm in the same case. The scenario was described in Section4, and we organized our network deployment in a similar way asClayton et al. [8] did. One sensor node was considered in each of the same places where the sensors of the related work were de-ployed, and additionally, two clusters were set up. Cluster j = 1comprising oors 1 and 2 (sensors i = 1 and i = 2, respectively),and the other cluster j = 2, comprising oors 4 and 5 (sensors i = 3and i = 4, respectively). The CHs j = 1 and 2 were considered out of thestructure, butstill in theradiorange ofallthe sensors.Thephys-ical presence of the CHs over the structure could affect the proper-ties of thestructure, adding more mass to it in a real situation. Sinceour CHs do not sense, our system could adapt very closely to theexperimental scenario described, as thesystemof Clayton et al.did.

Since there was information available about the healthy anddamaged states of the structures, it was possible to calculate D x i,t for t = 1 to5,and weassumedthatall the sensors detectedthe samevalues of modal frequencies. Since the structure is small and thesensors were relatively close, the sensing of the same global modalfrequencies of the structure by all the sensors deployed over it is areasonable assumption, but it is not true for every structure. Therewas information available to perform at least one rst (healthy)sensing andve successive data collectionstages with thepresenceof damage on each of the ve oors. Therefore, a study prior to theactual deployment of theWSNover thestructure was conducted fornding good values for the constants T i, Ai and L j.

In the experiment described in [8], three different methods forextracting the natural frequencies were used: a manual method,

the Eigen system Realization Algorithm (ERA) and the FractionalPolynomial Curve-tting (FPCF). The determination of the naturalfrequencies therefore presented an average and standard deviationamong theextractions through these three methods, for everysitu-ationof damage. The T i values assumed, foreverysensor i, the valueof 3 times the standard deviation of the respective modes of vibra-tion. Assuming that thenatural frequenciespresenteda normal dis-tribution, and considering that the algorithm uses the modulus of the difference of the modal frequencies, i.e., value of 3 correspondsto theamplitude of 6 times of thestandarddeviation, thealgorithmwill consider the structure healthy more than 99% of the times itsamples the natural frequencies. Only when a value of D x i,t sur-passes thevaluesof T i the algorithmproceeds withthe calculations,what corresponds (i) to 1% of the times even when the structure is

still healthy (false-positive) and (ii) to the situations where a realdamage has occurred. Once the changes in natural frequencies

due toa damage insertion ona oor in the structure were muchlar-ger than3 standarddeviations of thenatural frequenciesextraction,it was possible to successfully apply the algorithm. The value of L jwas dened in using the same principles, but according to the val-ues of the C j,t coefcient.

Once it was possible to previously calculate the values of D x i,t for every case of damage and every sensor, the determination of Ai

constants was made using a simple procedure. The A

i constantsshould be the ones which maximize the values of the expectedDi,t coefcients for thesensors localizedclose to each damagedsite.So, we solved a simple linear programming problem whose objec-tive was to maximize ( D1,1 +D2,2 +D3,4 +D4,5), varying the values of Ai and restricting them between 0% and 100% for each mode of vibration, for each sensor. Considering that in t = 1 damage waspresent in oor 1, in t = 2 in oor 2, t = 4 in oor 4 and t = 5 in oor5. So, we are maximizing the value of D i,t for the sensor i = 1 whenthe damage is in the oor 1, and so on. Since the values of D x i,t were the same collected by all sensors, they were constants inthe problem. Obviously, the modes of vibration which were moresensitive to the damage in one oor received 100% for the respec-tive value of Ai in the sensor of that oor. In other words, whendamage is at oor 1, and we know that the largest variation inthe frequency of mode 2 occurs when damage is at oor 1, wecan choose the value of 100% for this mode and 0% for all the othermodes for the Ai constant of the sensor at oor 1 ( i = 1). When thedamage is not at oor 1, this sensor may still detect a variation inthemodal frequency 2, andconsequentlymultiply this variation by100% generating some value of Di,t , but this value is still not thelargest value possible for Di,t .

Using these rules for setting the values of the constants of thealgorithm and considering the existence of two clusters, as previ-ously mentioned, allowed a reasonable performance of the algo-rithm. When damage was at oor 1 and at oor 2, cluster 1signalized damage near it, and cluster 2 did not. When damagewas at oor 3, since there was no sensor deployed onthis oor, thissituation was not predicted by our strategy, therefore both clustersacted in anunpredictedway, havingboth signalizeddamage.Whendamage was at oors4 and 5, theexpected behaviorwas for cluster2 to signalize damage alone, but since the damage localization onthese oors relied on the modal frequencies which mostly varieddue to damage in every oor, both clusters signalized damage inboth cases, what can be considered the case of a false-positive.Therefore, our algorithm presented 100% success in damage detec-tion and 2 cases of false positives among the 8 available results fordamage localization (two clusters signalizing damage or not nearthem in four predicted scenarios of damage). Although the perfor-manceof the algorithmcan be improved through choosingdifferentrules for setting the values of theconstants, it is extremelydepend-able of the quality of the structural properties. In cases where themonitored structuredoes not exhibit a clear standardof modal fre-

quency variations for each case of damage, it may be impossible tolocalize damage in some its places. In the scenario of Clayton et al.,damagesin oors4 and5 were not possibleto clearly localizeusingthe proposed rules.

6.9. Comparative analysis

We chose two important related works to perform a compari-son of our obtained results. The rst one is the work of Xu et al.[9], where the system named Wisden is introduced. This systemis considered a centralized approach, where no processing aimedat assessing the structural integrity is performed within the net-work, only at the sink node. Wisden is based on a WSN, but itsdesign is closer to the design of centralized wired systems. Wisden

relies on event detection and data compression techniques forproviding viable data transportation to the sink node, where a



14/16

complete analysis can be performed. The second work selected forour comparative analysis is described in Hackmann et al. [14],where a partially decentralized system for detecting and localizingdamage is presented. Such systemrelies on the DLAC algorithmfordetecting andlocalizing damage on structures, and is also based ona WSN. Part of the processing aimed at providing the inputs for theDLAC algorithm is performed within the network, but the DLAC

algorithm itself runs on the sink node. Both works are consideredprevious stages of the evolution in the use of WSNs for SHM, andpresent increasing levels of decentralization. Our work is, there-fore, intended to go a step further in this evolution, presenting afully decentralized approach. This comparison is designed to pointin which aspects the present work evolved with respect to a fullycentralized and a partially decentralized approach.

In Table 3 we compare our results in terms of (i) RAM memoryconsumption for performing the core functionalities which eachsystem proposes as solution for data reduction, in order to enableefcient transmissions, (ii) latency, which corresponds to the timedelayrequired for achieving theresults over a performed sampling,and (iii) network trafc generated by all the nodes of the network.In order to provide a fair baselinefor comparison, some data hadtobe normalized.

Wisden uses 288 bytes of RAM to perform compression over a128-sample array. Our solution consumes almost the doubleamount (considering the implementation of the Sensor node only),since it performs more costly procedures. Wisden uses a wavelet-based compression, while our system relies on a FFT and a peakextraction algorithm. Our proposal is more easily comparable toWisden, since the proposal of Hackmann et al. is based on theImote2 platform, which provides and consumes more computa-tional resources. However, Wisden presents a memory usage opti-mization, performing dynamic memory allocation, while oursystem consumes the same amount of RAM during its whole exe-cution, as the proposal of Hackmann et al. also does. The criterionof memory usage clearly points that our proposal performs a largeramount of decentralized processing than a fully centralized case.The work of Hackmann et al. points out an even greater amountof RAM consumption for each collected sample dealt with, but thisvalue is not fully comparable to our proposal because of two facts:(i) the Imote2 tool chain drastically inates the memory consump-tion of a program adding more software resources to it thanneeded, and (ii) the implementation of Hackmann et al. keepsthe arrays used in intermediary operations in the program fordebugging reas

a localized algorithm for structural health monitoring using wireless

Documents