An Integrated Health Management and PrognosticTechnology for Composite Airframe Structures

Ingolf Mueller 1, Cecilia Larrosa 1, Surajit Roy 1, Amrita Mittal 1

Kuldeep Lonkar 1, and Fu-Kuo Chang 1

1 Department of Aeronautics and Astronautics, Stanford University,Durand Building, 496 Lomita Mall, Stanford, CA 94305, USA

{imueller, clarrosa, fkchang}@stanford.edu

ABSTRACT

Next generation technology of integrated healthmanagement systems for air-transportation struc-tures will combine different single SHM methodsto an overall system with multiple abilities con-sidering different stages of damage initiation andpropagation. The fundamental configuration ofthe proposed SHM technique will involve the ideaof an integrated passive/active monitoring and di-agnostic system extended by numerical modulesfor lifetime prediction. The overall system is ca-pable of providing real-time load monitoring anddamage estimation on a global structure level aswell as precise damage diagnostics on a locallevel. This robust diagnostic technique providesquantifiable damage location and size estimationthat account for the uncertainties induced by theenvironments or the system itself continuouslyduring flight. Additionally, efficient predictionand prognostic methods are integrated with moni-toring and diagnostic outputs to provide real timeestimation of possible damage scenarios, residualstrength, and remaining useful life of the dam-aged structure. From this result information isgained which allow appropriate preventative ac-tions on the monitored structure. To achieve thoseobjectives, a built-in sensor/actuator network isemployed and numerical simulation methods ofdamage estimation and propagation are devel-oped and applied. The goal of this work is to inte-grate all these single techniques and subsystemsinto an integrated structural health managementsystem for composite airframe structures. Thesystem design, data exchange between the dif-ferent subsystems, and the performance of eachmodule is presented.

This is an open-access article distributed under the termsof the Creative Commons Attribution 3.0 United States Li-cense, which permits unrestricted use, distribution, and re-production in any medium, provided the original author andsource are credited.

1 INTRODUCTION

Structural design of air transportation systems includesmany goals: light weight for fuel efficiency, stiffnessand strength for an appropriate dynamic performance.Another major task is the improvement of the safetyof those systems by inspecting for system and com-ponent failure. At the same time the downtime of thetransportation system due to inspection should be dra-matically reduced by nearly continuous monitoring ofthe structure.One of the most critical concerns is the presence of un-detectable damage, which may be introduced duringthe manufacturing process as well as in-service events,such as impact loads or fatigue. The presence of anydegradation, which is not easily detectable by visualinspection, may significantly reduce the mechanicalproperties of a component, in particular strength andstiffness. Due to this fact, the design of air transporta-tion systems and associated maintenance practice musttake this concern seriously into consideration in prac-tice.Due to the increasingly high demands for safety andlow cost maintenance, the use of a built-in real-timestructural health monitoring (SHM) system for aircrafthealth monitoring is becoming more and more attrac-tive. Unlike traditional NDE systems, the SHM systemis designed to apply to a specific structure with a built-in network of sensors and actuators.Although many well-established diagnostic techniquesand prognostic methods have been proposed in the lit-erature during the last decade, each single technologyof structural health monitoring aims at observing veryspecific properties of structures or systems. Due to thisfact, the performance of a single technology is sub-jected to inherent limitations of the applied method-ology and may only cover single aspects of structuralhealth monitoring. In an effort to improve the over-all system performance and to merge interactively thecapabilities and advantages of the several single meth-ods, research effort is conducted to link and to inte-grate different SHM technologies to a complete sys-tem for the purpose of an Integrated Vehicle HealthManagement (IVHM) as required for next-generationair and space transportation systems. The goal of

1

Figure 1: Flowchart of the interactive health management system as proposed.

IVHM is to develop validated technologies and tech-niques for automated detection, diagnosis and prog-nosis to enable early detection and mitigation of ad-verse events during flight. Adverse events includethose that arise from system or component faults orfailures due to damage, degradation, or environmen-tal hazards that occur during service of vehicles. Thecapabilities of the IVHM systems will comprise rapiddetection and diagnosis of adverse events, estimationof severity and remaining useful life (RUL) for the af-fected systems. The gained health information will bedistributed to other subsystems as real-time automatedreasoning and decision making systems. The integra-tion of monitoring and diagnostics with damage pre-diction and prognostics into a health management sys-tem is still a very challenging technical task and thesubject of the current research. In particular, the ex-change and the transfer of health data between the dif-ferent subsystems as well as an efficient interaction toachieve optimal performance need to be explored indetail and receive immerged attention. The details ofdata flow and linkage between the subsystems are il-lustrated by the flowchart of Figure 1.After final completion, the proposed technology willbe capable to diagnose automatically in real-time thehealth condition of a structure as well as provide anestimate of the residual strength and remaining usefullifetime to optimize the performance and off-serviceschedule of the transportation system. Accordingly,the health management system relies on a network ofsensors and actuators to be integrated with the struc-ture and be equipped with a diagnostic capability fordetecting damage and a prognostic capability to pre-dict the residual life of the damaged structures in ser-vice.By merging all the aforementioned single techniquesand subsystems an integrated structural health man-agement system for composite airframe structuresis obtained having unique capabilities of interactivemonitoring, detection and prediction.

After a short introduction of the embedded networktechnology, each major SHM subsystem is briefly de-scribed along with interfaces to the linked subsystems.The current state of work and open questions are dis-cussed.

2 PASSIVE AND ACTIVE MODE OFEMBEDDED SENSING NETWORKS

Structural health monitoring techniques require the useof built-in sensors in a network to monitor externalconditions and structural integrity. In general, thereare two types of sensor-based diagnostic and moni-toring systems: passive sensing (e.g. (Doyle, 1987;Yen et al., 1994; Jiang et al., 1998; Fogg et al., 1993;Seydel and Chang, 1999; Chang and Seydel, 2001;Park, 2005; Markmiller, 2007)) and active sensing(e.g. (Chang and Beard, 1997; Chang and Roh, 1995;Chang and Ihn, 2004)). In the majority of cases, piezo-electric sensors are employed to build sensing net-works. Piezoelectric sensors are active transducers,which can act for both the collection of measurementdata at certain points of the structure and the genera-tion of controlled diagnostic signals.A passive system is useful to monitor changes in theenvironment, such as loads and impacts, but it is ex-tremely difficult to locate damage or cracks with highreliability and good resolution. However, innovativepassive technologies may have the capability to pro-vide information on a global structure level about theexistence and rough location of damage, which mightbe considered within a global condition monitoringprocess. Since no actuators are needed here (only op-erational excitation is used) a limited hardware is suf-ficient to operate the system. For the proposed healthmanagement system, the passive network technologywill mainly be used to acquire real-time sensor datafor impact monitoring as one crucial SHM subsystem.

Re-examination and interrogation in order to locateand characterize damage, as well as to estimate theresidual life of structures, becomes possible within an

2

Figure 2: Capabilities and flow of structural health data of the interactive health management system.

Figure 3: Passive (left) and active sensing networks(right).

active sensing approach. In principle, an active sys-tem allows damage to be interrogated by injecting con-trolled diagnostic signals (i.e. guided Lamb waves)into the structure. With the known inputs, the changesin local sensor measurements are associated with theintroduction of damage in the structure. Here, the ac-tive sensing approach compares the sensor data beforeand after an event to interrogate the structural condi-tion. Therefore, it is most effective to detect dam-age and local defects in structures. Although thesetechniques have been demonstrated to be a promisingmethod for reliable damage diagnostics, active sys-tems have the disadvantage that extensive hardwareand power supply is needed to operate the system.Additionally, the installation of active networks maynot be feasible on the entire structure and usually onlymonitoring of structural hot-spots is performed.Current research, as described in this contribution,aims to link the two sensing technologies in an effec-tive way which allows to merge the capabilities andadvantages of the two sensing technologies. The pas-sive technology constantly monitors the structure andassociated algorithms will provide real-time capabil-ities (Park, 2005), (Markmiller, 2007) for global im-pact and condition monitoring. After detection of a se-vere impact event, data of impact location and impactforce history will be submitted to a damage predictionmodule where numerical simulation is used to estimatethe impact damage by failure analysis. Additionally,

the active technology provides the damage diagnosticsand will afterwards interface with numerical prognosismodule for estimation of remaining useful lifetime.In summary, the uniqueness of this approach is an in-teractive global-local diagnostics by integration of pas-sive sensor data as well as active sensor data providedby locally built-in actuators in the structure. The over-all system will be capable of providing real-time loadmonitoring on a global structure level as well as pre-cise damage diagnostics on a local level. For this pur-pose, structural health data will be exchanged betweenthe subsystems and numerical models will be continu-ously updated with occurring degradation.

3 IMPACT MONITORING BY PASSIVESENSING NETWORKS

Real-time impact monitoring is an important key ap-proach for structural health management. For this pur-pose, a passive sensing network based on built-in sen-sor technology is employed to acquire local sensor dataof structural response due to impact events. However,passive sensing approaches rely on numerical modelsto reconstruct meaningful information as impact loca-tion and impact load history from the point-wise sen-sor data.Generally speaking, recovering external impact loadsusing structural response data works on the principleof an inverse problem (see Figure 4). While the phe-nomenon of interest (impact location and force his-tory) may not be measured directly, there exists someother variable that can be observed (sensor data ofstructural response), which is related to the event ofinterest through the use of a data-derived computermodel. Depending on its mathematical representa-tion, the properties of the adopted model, and thesensor network configuration (number and location ofsensors) the resulting inverse problem may be sub-jected to the difficulties of ill-posedness. Accordingto Hadamard (Hadamard, 1923), the ill-posed proper-ties of the resulting inverse problem appear in termsof stability, existence, and uniqueness of the solutionand may cause serious errors in the identification re-

3

sults. The effect of ill-posedness for the problem ofimpact monitoring has been studied in (Mueller andChang, 2009), in particular with a view to influence ofthe sensor network configuration.

Figure 4: Schematic of inverse problems.

Several models and techniques (e.g. (Doyle, 1987;Yen et al., 1994; Jiang et al., 1998; Fogg et al., 1993;Seydel and Chang, 1999; Chang and Seydel, 2001;Park, 2005; Markmiller, 2007)) have been proposedin the literature to identify impact events from passivesensor data. Park and Chang (Park, 2005) have pro-posed an advantageous methodology which is basedon a linear system identification model. This auto-regressive model with exogenous input (ARX) (Ljung,1999) is a linear finite difference model which repre-sents the dynamic response of a linear elastic structureby a parametric approach as given by equation 1.

ε(k) =n∑

i=1

ai ε(k − i)+m∑

j=1

bj f(k − j) (1)

Here, ε(k) and f(k) are a sequence of strain and im-pact force values at discrete time instances. The pa-rameters ai and bj are the ARX parameters of themodel while n and m represent the order of the model,respectively. A training step is needed to find appropri-ate ARX parameters based on the physical behavior ofthe structure. For this purpose, sets of training data interms of impulse response are generated and ARX pa-rameters are derived.For the inverse reconstruction of the impact force his-tory the ARX model need to be inverted. One strik-ing feature of ARX based methods is that the inversioncan be performed analytically as shown by Park (Park,2005). Due to this fact, real-time capabilities of im-pact monitoring module are achieved. Furthermore,this method can be applied to complex structures withuncertain properties and boundary conditions. On theother hand, the proposed ARX model has inherent as-sumptions of linearity. Non-linear structural behav-ior as occurring on time-variant structures (e.g. dam-age propagation) may not be represented correctly.For these cases, a model update or the utilization ofanother set of ARX parameters might be necessaryto capture appropriately the current condition of thestructure. Hence, a feed-back of the current health sta-tus received from active diagnostics is needed to in-volve the latest structural condition.

Figure 5 illustrates the results of the impact monitor-ing system on the example of a stiffener panel manu-factured by Boeing Company (also see Figure 14). Theproposed method identifies the impact location as wellas the impact load history in real-time. As a compari-son, the impact force was directly measured by utiliza-tion of a modal hammer (red curve) and reconstructed

Figure 5: Results of impact monitoring on the exampleof a composite stiffener panel (left: impact locationand energy map, right: impact force history).

by the described ARX approach (blue curve). As canbe seen, the results are in a good agreement.

3.1 Stability of ARX Models for ImpactMonitoring

One significant difficulty in parametric identificationof ARX models is an appropriate assumption for themodel order n (see Equation 1) to construct the regres-sion vector. In fact, an incorrect choice for the modelorder n will lead, in general, to difficulties in terms ofmodel accuracy.

Figure 6: Stability map of ARX models for differentmodel orders n (upper: n = 3, lower: n = 5).

4

Selecting a value for n too large is called over-parametrization and may lead to unstable ARX esti-mates in the majority of cases. On the other hand, se-lecting a value n too small will lead to mismatch be-tween measured data and model output. In this case,errors of approximation may be large. A heuristic ap-proach for estimation of an appropriate model ordercan be found in (Park, 2005).In order to evaluate the model order for a given struc-ture a stability map for the estimated ARX parametercan be compiled. It can be expected that almost iden-tical or similar impact events will only lead to slightchanges for the estimated ARX parameters. Sucha stable behavior can only be found for appropriatemodel orders n. In all other cases, the model mightbe subjected to instability and, hence, slight changesin the input/output data leads to large differences inthe computed ARX parameters (fluctuations).Practically, a large number of input/output data will berecorded based on impacts at the same location andwith similar amplitudes (impact force). For each setof I/O data an estimation of ARX parameters is per-formed and results are collected in a stability map (seeFigure 6). For stable model orders n straight lines forthe ARX parameter are obtained. Figure 6 shows atypical result on the example of the before-mentionedcomposite stiffener panel. A model order n = 3 (up-per) will show stable ARX models while a 5th or-der model (lower) leads to strong parameter fluctua-tions. As a result of extensive experimental tests, ithas turned out that model order n = 3 is the most ap-propriate assumption for the considered class of struc-tures.

3.2 Automated System Calibration andSimulation-Based Training of ARX Models

In order to improve and to speed up the necessarytraining step for the ARX model Markmiller andChang (Markmiller, 2007) recently proposed the uti-lization of a structural model based on Finite Elementmethods to obtain numerically the necessary impulseresponses of the structure. Figure 7 shows the two-stage approach in a schematic.

Figure 7: Training of ARX models by numerical sim-ulation.

Apparently, the quality of the training processstrongly depends on the accuracy of the Finite El-ement model employed during the training process.Stated another way, without having a reliable simula-tion model at hand, the training step of finding appro-priate ARX parameters may lead to incorrect results.Thus, the calibration of the numerical model and thesensor / model relations is of immense importance.The calibration of the sensor/model relation involves

usually extensive experimental testing which mightbe difficult, time-consuming, and expensive for large-scale structures. In particular with a view to futurelarge sensor networks – having potentially thousandsof sensors in a network – a reliable and efficient cali-bration technique is strongly required.A promising alternative to extensive experimental labtests is to extract the necessary dynamic properties forthe numerical model from the structure response col-lected by the passive network during operation (oper-ational data / in-flight data) or noise tests (Chang etal., 2008). Here, the challenge appears in the reliableextraction of needed system properties (system iden-tification) from sensor data only (dynamic response)while the ambient input (dynamic excitation) is not ornot completely known. In contrast, commonly appliedprocedures of system identification are based on in-put/output relations. Nevertheless, the key target ofboth approaches is the same, namely to get knowledgeeither of frequency response or impulse response func-tions of the system under consideration. The obtaineddata are either used to update the numerical model interms of model calibration or provide a direct basis forthe training process.

Figure 8: Extracted vs. analytical frequency responsefor the example of a beam-type structure (upper: after200 sec measuring time, lower: after 2000 sec measur-ing time).

5

In general, the methodology of output only systemanalysis - as core for automated sensor / model cali-bration - is based on three foundations: (1) long-termmeasurements with appropriate sampling rate by thepassive sensing, (2) averaging in order to reduce mea-suring noise and random effects, and (3) power spec-tral density estimate. Following the assumption thatthe natural (operational) excitation of a flexible struc-ture is characterized by continuous energy supply witha broad spectrum, one can expect energy to concen-trate around modal frequencies. Hence, a power spec-tral density estimate will lead to the according fre-quency response function that can be utilized for a sub-sequent model updating process or direct training step.Figure 8 shows an example for the reconstruction offrequency response functions by output-only analysisas part of the proposed calibration method.

4 DAMAGE PREDICTION AFTER IMPACTDETECTION BY PASSIVE SENSINGNETWORKS

To accomplish the goal of a health management systemwith immediate capabilities for damage prediction af-ter an adverse impact event, the results of impact mon-itoring will be linked with numerical failure analysis.Once the impact load history is known by the impactmonitoring system, an appropriate FE model for thestructure is employed to obtain the stresses during theimpact event. The damage will then be estimated fromthe stresses using appropriate failure criteria. In thisresearch, criteria for matrix cracking and delaminationhave been considered and implemented.

A lot of research has been conducted during the pastdecades to investigate the type, location and the extentof damage in the fiber reinforced laminated compositesdue to impact events. As commonly acknowledged,transverse low velocity impact on laminated com-posites mainly induces intra-ply matrix cracking andinter-ply delaminations (Choi, 1990; Wu and Springer,1988; Wu and Chang, 1989; Geubelle and Baylor,1988). The occurrence of cracks is primarily due tothe inter-laminar transverse shear stress and transversein-plane stress. Delamination growth at the interfaceis governed by the state of stresses of the constrain-ing plies or ply groups having different orientations.Many researchers have developed analytical models toestimate the type and extent of damage due to the lowvelocity impact forces. A detailed investigation onthe actual modes and mechanisms of impact induceddamage on the laminated composites was presented byChoi (Choi, 1990).The analytical model for predicting the delaminationin the composite structures comprises first carrying outa transient dynamic analysis under the application ofthe load history obtained from the integrated passivesystem and subsequently applying appropriate failurecriteria (Choi, 1990), (Hashin, 1980) to capture theinitiation sites and the extent of damage. The tran-sient dynamic analysis of a composite structure is per-formed using the commercial Finite Element packageABAQUS along with a UMAT user-defined materialsubroutine. Once the stresses are obtained the follow-ing criterion is employed to predict the occurrence of

matrix cracking(

σyyYt

)2+

(σyzSi

)2≥ 1 (2)

where x, y and z are the principal material directionsof the laminate. Here, Yt is the in-situ transverse ten-sile strength and Si is the in-situ ply transverse shearstrength. Both these strength parameters depend on therelative thickness and the orientation of the constrain-ing ply-groups. These strength parameters are deter-mined from the empirical relationship as found in theliterature (Lessard, 1989).

Figure 9: Flowchart for predicting the extent of delam-ination under a given impact load.

Once the matrix failure criteria is satisfied at a par-ticular material point in any ply group near the inter-face of the two different ply-groups (having differentmaterial orientations) then for all the subsequent timesteps of the dynamic analysis the possible occurrenceof delamination and its growth is monitored. The fol-lowing criterion (Choi, 1990) is used to estimate thepresence of delamination

Da

[〈n〉

(σyzS

)2+〈n+1〉

(σxzS

)2+〈n+1〉

(σyyY

)2]≥ 1(3)

where n denotes the upper layer and n + 1 the lowerlayer of the interface. The stresses σyz and σxz repre-sent inter-laminar shear stresses while σyy is the trans-verse stress in a ply. Again, Y denotes the transverse

6

strength and S represents the shear strength. The fac-tor Da is an empirical constant which has to be deter-mined by experiments. For graphite epoxy compositesT300/976 an average value of 1.8 has been found inliterature. The flowchart of Figure 9 depicts the algo-rithm which has been implemented in the user-definedsubroutine UMAT.The validation of the UMAT material subroutinewhich has been developed to predict the location andthe extent of the delamination in a ply under a givenimpact load history is carried out with the experimen-tal results available in the literature (Choi, 1990).

Figure 10: Geometry of the flat plate considered forvalidation of the implemented UMAT subroutine.

A flat composite laminated plate 3.94” × 3.00” ×0.09”, clamped at the two outer edges was impactedwith spherical projectiles directed at the center. A lin-ear 8-noded 3d composite solid brick element (C3D8)is used to model the structure with one integrationpoint along the thickness direction for each ply in thelay-up. The material properties of the plate are shownin the table below (Choi, 1990). The material of thespherical projectile is taken as steel with a mass of0.353 lb.

Table 1: Material properties for ’T300-976’ as avail-able in the literature.

Material Property Symbol Value[Unit]

a) Elastic Constantsin-plane longitud. modulus Exx[psi] 0.224 E08in-plane transverse modulus Eyy[psi] 0.130 E07in-plane shear modulus Gxy [psi] 0.995 E06out-of-plane shear modulus Gyz[psi] 0.463 E06in-plane Poisson’s ratio νxy [psi] 0.228out-of-plane Poisson’s ratio νyz[psi] 0.400b) InertiaDensity �[lbm/in3] 0.056c) Strengthlongitudinal tension XT [psi] 0.217 E06longitudinal compression XC [psi] 0.227 E06transverse tension YT [psi] 6435.0transverse compression YC [psi] 0.360 E05ply longitudinal shear S[psi] 0.145 E05

First, a convergence study for the Finite Elementmodel has been carried out to understand the effect ofthe element size on the plate’s central deflection as de-picted by Figure 11. In the following a total of 900 el-ements is selected for the analysis as result of the meshconvergence study.

Figure 11: Mesh convergence study for the maximumcentral deflection of the plate under the specified im-pact force distribution.

Figure 12 shows the comparison between the pre-dicted and the observed delamination sizes along theinterface that has the maximum delamination size inthe composite lay-up for different projectile velocities.For a [−454/458/−454] lay-up (Figure 12, top) themaximum delamination size is between the 4th and 5thinterface when counted from the bottom ply. The loadis applied at the center of the top ply which is definedas the 16th ply. Similarly, for a [04/452/−454/452/04]lay-up (Figure 12, bottom) the maximum delaminationoccurs along the 4th and 5th interface located between00 and 450 ply.

Figure 12: Comparison of the delamination sizebetween the experimental values and predictionsfrom the numerical simulation for T300/976 com-posite (upper: [−454/458/−454] lay-up, lower:[04/452/−454/452/04] lay-up).

It can be observed from Figures 12 that the predictedvalues match well with the experimental observationsfor the delamination sizes. The method thus developedcan be used to predict the delamination in a composite

7

lay-up once the loading history is known from passiveload identification.Figure 13 (top) shows the force history as obtainedfrom a real impact event monitored by the passive sys-tem. Following, this impact force history is appliedto the failure analysis model in order to predict thedamage in terms of delamination. As an example, aflat composite plate 3.0” × 3.0” × 0.1” clamped atall the edges has been considered and the lay-up be-ing assumed as [0/−45/45/90]2S. It is observed fromthe simulations that no damage will occur along anyinterface for the given loading conditions originallyderived from an integrated passive monitoring system.Accordingly, the load history is linearly scaled

f�impt(t) = α · fimpt(t) (4)to predict the threshold value of the peak load whichwill cause the first delamination in any one of the pliesof the laminated structure. It is found that the delam-ination will occur only when the peak load value isincreased by factor α = 3.

Figure 13: Impact force history as obtained fromthe integrated passive system (upper); extent of de-lamination at the 3rd interface for the T300-934[0/−45/45/90]2S composite (lower); note: the impactforce history is scaled by factor α = 5 with respect tothe shown curve as obtained from the integrated pas-sive system (upper) to produce the extent of damage asshown.

Once the impact force distribution on a laminatedstructure is known, the state of damage in the structurein terms of delamination can be computed. In addi-tion, an estimate of the peak load required to cause thefirst occurrence of damage in the laminated structurecan be predicted. Figure 13 (bottom) depicts the extentof delamination between 2nd and 3rd interface for theT300-934 [0/−45/45/90]2S composite when the peakload is increased by a factor of α = 5.

In summary, it has been demonstrated that diagnos-tic load monitoring by passive sensing data with alink to impact damage prediction has great potentialfor the development of next-generation health man-agement systems. The state of delamination (locationand size) in any composite laminate can be evaluatedby using the developed material subroutine UMAT inconjunction with the commercial Finite Element pack-age ABAQUS. Future work will be conducted to uti-lize the damage information thus available to providea first estimate for the health of the structure in termsof the residual strengths and residual fatigue life.

5 ACTIVE SENSING FOR DAMAGEDIAGNOSIS

One of the major advantages sensing networks basedon piezoelectric transducers is that it can be used foran active and passive mode. In the active mode, eachpiezoelectric element in the sensor network can beused as both a transmitter and a receiver so that morecomprehensive information about the structure and thedamage can be retrieved.The fundamental principle of damage diagnosis by ac-tive sensing is based on the controlled injection of di-agnostic waves into the structure under observation. Itis widely acknowledged that the utilization of guidedLamb waves is one of the most promising approachesin this area which is increasingly applied in variousengineering fields. Some of the fundamental advan-tages over other methodologies are the capabilities toinspect large structures and the entire cross-sectionarea without disassembling the system under inspec-tion. A comprehensive review of the current state forthe guided Lamb wave technique can be found in (Suet al., 2006).

The acousto-ultrasound (AU) technique based onLamb waves has been demonstrated to be very sensi-tive and effective in detecting of local damage (Wangand Chang, 2000), (Park and Chang, 2003), (Kim etal., 2008), (Boller et al., 2009), in thin to moderatethick structures which is ideally for aircraft structureapplication. Advanced diagnostic images similar to ul-trasonic images have been developed to identify the lo-cation of damage and to estimate its extent.Although, the damage diagnosis by active sensing hasattracted much research effort during the last years andextensive studies have been conducted, still some fun-damental difficulties and open questions exist that re-quire pursuing the research efforts. In the following,a question with particular relevance for the proposedinteractive health management system has been takenout and will be discussed.

5.1 Compensation of Environmental EffectsAn active diagnosis system allows damage to be inter-rogated by injecting controlled diagnostic signals intothe structure. With the known inputs, the changes inlocal sensor measurements (compared to baselines ofa pristine state) are associated with the introduction ofdamage in the structure.In general, a set of baseline data will be collected fromall actuator / sensor paths on the intact structure. Thesignals will change if any local degradation occurs. Toquantify the change in the signals due to damage, the

8

post-damage signal can be subtracted from the origi-nal (baseline) signal before there was damage. The re-maining difference is called scatter signal that is usedto quantify the amount of change in the signal of a par-ticular actuator-sensor path. After data normalization(e.g. taking path length into account) of the scatter sig-nal amplitude, the ratio between signal scatter and sig-nal baseline within a certain time window [t0, t1] pro-vides an indicator for structural alteration, as shown byEquation 5.

SER =

⎡⎢⎢⎢⎢⎢⎢⎢⎣

t1∫t0

‖SSC(ω0, t)‖dt

t1∫t0

‖SB(ω0, t)‖dt

⎤⎥⎥⎥⎥⎥⎥⎥⎦

n

(5)

Compiling this information for each sensing path cre-ates the basis for the diagnostic image. Diagnosticimaging which highlights the region where signifi-cant changes in sensor-actuator signals occur has beendemonstrated to be a beneficial tool for locating dam-age and potentially quantify the damage.

Figure 14: Example of damage diagnosis by guidedlamb waves on the example of a stiffener compositepanel with generic damage.

However, the acousto-ultrasound technique is notonly sensitive to structural alterations as induced bydamage, it is also sensitive to the change of envi-ronmental condition (Scalea and Salamone, 2009),(Raghavan and Cesnik, 2008). Structures with inte-grated health monitoring systems operate across a va-riety of environmental conditions including changes intemperature, pressure, humidity, static load and vibra-tion. The effect of these factors on Lamb wave prop-agation can easily be interpreted as an indication ofstructural damage or, stated another way, environmen-tal effects can corrupt the sensor signals in such a waythat an accurate prediction of location and size of thedamage will not be possible in the majority of cases.Sensor signals change because variation in environ-ment alters material properties of the host structure,

PZTs, and the bondline adhesive layer. Hence, in or-der to avoid false warnings, the SHM technology mustbe compensated for environmental factors for fieldedapplication.

Figure 15: Example for the influence of different en-vironment temperatures to the scatter energy ratio fordifferent scanning frequencies (100 kHz ... 400 kHz).

Current approaches try to overcome these difficul-ties by utilization of large sets of different baseline datafor different environmental conditions to compensateenvironmental effects on damage diagnosis. More-over, interpolation techniques are applied to providebaseline data for all possible conditions. However,interpolation in heterogeneous data may be a sourcefor severe errors in damage diagnosis, in particular, ifbased on a small number of experimental sets of base-line data. On the other hand, the acquisition of a largedatabase for the necessary baseline data might be time-consuming and inefficient even for large structures.

In an effort to improve the current compensationtechnique research is conducted to develop a methodwith minimum of necessary baseline data to compen-sate for the environmental effects on sensor signal.The method to be developed will be based on thephysics that causes the environment-induced changesin the diagnostics signals. For this purpose, the physicsof wave propagation under certain environmental con-ditions need to be studied and simulated by numericalanalysis.However, direct relationships between the signalchanges and environmental conditions are still quitedifficult to establish, and such a relationship maychange with sensor/actuator location and structural ge-ometry, etc., making it difficult to quantify and to com-pensate the environmental effects in a real application.Currently, numerical tools based on Finite Elementmethods are developed to achieve reliable simulationfor the wave propagation of Lamb waves in struc-tures of practical complexity. The simulation is basedon the application of spectral elements (Komatitschand Vilotte, 1998) within an explicit time integrationscheme. This methodology was presented recently byChang and his associates in (Kim et al., 2008) andturns out to be promising as a fundamental basis forenvironmental compensation (Ha et al., 2009).The following study is aimed at understanding the ef-

9

fect of elevated environmental temperature on Lambwave propagation. Experiments are carried out overtemperature range from 250C to 750C where a signif-icant change in sensor signal with temperature is ob-served. A parametric study is performed by simula-tion to identify the key parameters that change Lambwave propagation with temperature and the bondlineadhesive is found to be one of the most critical factors.

Figure 16: Experimental setup of an aluminum platewith surface mounted PZT transducers (all dimensionsare in mm).

An aluminum plate with 0.8 mm thickness is usedto investigate the effect of temperature. Dimensions ofthe plate are shown in Figure 16. Piezoelectric trans-ducers (PZT 5A) of 0.25 mm (10 mil) thickness andwith a quarter inch diameter were glued to the surfaceusing conductive epoxy (Chemtronics cw2400) at lo-cations indicated in Figure 16. The PZTs are gluedwith varying amount of glue to study whether adhe-sive thickness affects the Lamb wave propagation withtemperature. The specimen is heated in an oven to el-evated isothermal temperatures. A ScanGenie scan-ning device from Acellent Technologies is employedfor data acquisition. Here, a 5 peak tone burst wave hasbeen selected as actuation input each at 300 kHz andsensor signals are measured at 250C, 500C, and 750Cwith 24 MS/s sampling rate in a pitch-catch mode.The aforementioned Spectral Element Method (SEM)is employed to simulate the propagation of Lambwaves in the described aluminum plate with surfacemounted PZT actuator and sensors. For this purpose,0.04 mm and 0.10 mm thicknesses are assumed forthin and thick adhesive layers respectively.

In general, results of the numerical analysis showa dominant influence to the signal by temperature-induced property changes for the adhesive layer of sen-sor network. Exemplarily, Figure 17 compares simu-lation results in terms of time signals for Lamb wavepropagation. In order to illustrate the influence of tem-perature change to the diagnostic signals induced bychanges in material properties (Sherrit et al., 1999),(Brammer and Percival, 1970) the Lamb wave propa-gation is simulated for the described aluminum platesubjected to a temperature change of 75 K.For comprehensive parametric studies, material prop-erties are selected as parameters. Sensor signals aresimulated at 250C, 500C, and 750C. Properties are var-ied individually to study the effect of each parameter

Figure 17: Numerical simulation of Lamb wave prop-agation in a thin aluminum plate subjected to differ-ent temperatures as environmental effect (upper: ef-fect of changes in aluminum properties, lower: effectof changes in adhesive properties).

on the peak-to-peak amplitude and time of arrival ofthe sensor signal. The variation of adhesive stiffnesscan only be estimated and a 10 % reduction in stiff-ness of adhesive for 25 K rise in temperature has beenassumed for the following parametric studies.

Figure 18 shows the percentage change in peak-to-peak amplitude and time of arrival for 100, 300 and500 kHz excitation frequencies (for a signal travelingbetween transducers both mounted with thick adhe-sive). Percentages are calculated based on signals atroom temperature ( 250C). Positive percentage valuesindicate an increase in peak-to-peak amplitude or timeof arrival. From Figure 18 it can be concluded that ma-jor parameters affecting amplitude are adhesive shearmodulus (GAdh) and permittivity of PZT in thicknessdirection (ε33). In the other hand, the major influenceon time of arrival are given by the parameters of alu-minum stiffness (EAl), aluminum density (�Al), andadhesive shear modulus (GAdh).Figure 19 compiles experimental and simulation re-sults. As can be seen, for higher temperatures the am-plitude tends to decrease and delay of time arrival isobtained. Numerical simulations show good correla-tions with experiments in trends of both amplitude andarrival time. Discrepancies in experimental and simu-lation results are mainly due to uncertainty in variationof material properties with temperature.

In summary, the developed simulation technique isvery promising to understand the physics behind theenvironmental influence to the propagation of Lambwaves. The presented study has shown the effects oftemperature-dependent material properties to the wavepropagation as one important environmental factor. Inparticular, the dynamic behavior is affected by theproperties of adhesive layers, which has hardly beenstudied so far. The trends in experiments are simulatedwith SEM and explained based on the knowledge ac-quired from the parametric studies.

10

Figure 18: Parametric study on the influence of differ-ent temperature-induced parameter changes to prop-agation of Lamb waves (upper: peak-to-peak ampli-tude, lower: time of arrival (ToA)).

Figure 19: Comparison of experimental and numericalresults on the influence of temperature effects to Lambwave propagation (left: peak-to-peak amplitude, right:time of arrival, upper: simulation results, lower: ex-perimental results).

6 OUTLOOK – PROGNOSIS OF REMAININGUSEFUL LIFETIME

The current work focuses on the development of anappropriate technique for the prognosis of remaininguseful lifetime of structures based on the outputs ofdamage diagnosis and prediction of impact-induceddamage. In the following, the key steps of the pro-posed methodology will be primarily presented anddiscussed rather than extensive results. Furthermore,fundamental examples will provide basic validation ofthe procedure and will show the feasibility of the inte-grated lifetime prognosis.There is a significant difference between Health Mon-itoring and Health Management. Health Managementuses Health Monitoring to detect structural damage,then evaluates how the structure is affected by the dam-age (Prognostics – Residual strength) and estimateshow long the structure can operate after this damage(Prognostics – Remaining useful lifetime). Accord-ingly, the strategy of the integrated system comprisesthat after damage diagnosis by the active sensing sys-tem, the current health status of the structure is submit-ted to the prognostic module to achieve certain prog-nostic information. In order to yield accurate predic-tions of damage growth and residual strength, one keyissue is to create a proper link between the diagnos-tic outputs and prognostic inputs. This link can bebroken down into four main modules. Currently, theauthors are focusing on the diagnostic systems (as de-scribed above) and on the progressive failure analysismodule. The remaining diagnostics / prognostics (DP)link modules, described in detail below, constitute theroad map to be followed for the successful deploymentof the proposed health management system. Figure 20presents the flow of inputs and outputs within the dif-ferent modules.

A.) Health Data Transfer. The goal of this step is tomap the diagnosed damaged area to the structure’s Fi-nite Element model. The error in detection and extrac-tion of health data can significantly affect the accuracyof the life-time prediction. Therefore, the extraction ofhealth data from the diagnostic results is of immenseimportance.

B.) Identification of Damage Mechanism. Due toanisotropy of laminated composites there are differentdamage processes within the laminate: (1) TransverseMatrix-cracking, (2) Fiber-matrix debond (splitting),(3) Delamination, and (4) Fiber breakage. Dependingon the loading situation one type of fracture will nu-cleate and progress to the others. Characterizing thediagnostic outputs with failure modes is crucial and fu-ture research need to be conducted to overcome thesedifficulties. The current approach is as follows. Bycollecting the data of degradation by damage diagnosisand determining or predicting the loading spectra dur-ing operation an estimate of the sustained damage typecan be made. This running estimate of the damage ac-cumulation can be recorded and reconciled with futureresults of damage diagnosis at less frequent intervals.With these estimates either damage tolerance or safe-life philosophy can be applied, as appropriate for theparticular case, to estimate remaining useful lifetimes.

11

Figure 20: Flowchart of link from diagnostics to prognostics (DP link).

C.) Management / Decision. This module’s objec-tive is to provide the progressive damage analysis witha future loading scenario to estimate damage progres-sion. These future load spectra can be either damagetolerance or safe-life philosophy. The damage toler-ance approach assumes that a crack or flaw exists in apart and then determines the impact that a flight spec-trum will have on the growth of the crack size. Safe-life methodologies are based on extensive service his-tory and can determine the remaining lifetime of a partwhere no crack or flaw is expected to exist. By uti-lization of database techniques, an update of the RULestimate after occurrence of a specified trigger event(i.e. impact) can be achieved near real-time and willbe submitted to the decision-making module.

D.) Progressive Failure Analysis. The module ofProgressive Failure Analysis evaluates structure per-formance and calculates a remaining life-time predic-tion given the current state of damage and future load-ing. A progressive failure analysis model, originallyderived by Chang and his associates (e.g. (Changand Qing, 1997)) for the simulation of damage inbolted composite joints, was adopted and extended.A Hashin-based failure criterion (Hashin, 1980) wasselected to predict the damage during the evolutionof load history. When the criterion is fulfilled dam-age is introduced or propagated. It accounts for ma-trix micro-cracking, and uses degradation factors toaccount for other damage mechanisms, while it doesnot predict delamination. Future work will focus onextending the analysis to include delamination follow-ing the methodology described in the previous sec-tions. Upon damage the stiffness and strength of thelaminate are degraded following a degradation modeldeveloped based on fracture mechanics (Shahid andChang, 1993). The model was implemented as auser defined material constitutive model subroutine(UMAT) in the commercial Finite Element packageABAQUS/Standard.

The current modified version of the analysis accountsfor low cycle fatigue loads using a ply based frac-ture mechanics approach that predicts matrix micro-cracking initiation and propagation. As has been de-scribed above, matrix micro-cracking is the initialcause of delamination (Choi, 1990), and therefore,an accurate prediction of matrix cracking is necessaryprior to modeling delamination.First, the traditional approach of low cycle fatiguegrowth was implemented (Harris, 2003), (Nairm,1999) using Paris Law. In order to validate the al-gorithm, results of the numerical analysis were com-pared to published experimental data (Lafarie-Fernotet al., 2001), (Nairm, 1999). As can be seen fromFigure 23, the traditional approach results follow theexperimental trend. Still there are significant differ-ences between experiment results and numerical anal-ysis mainly caused by uncertainty in material proper-ties. Performing according experiments for materialcharacterization may be time consuming and costly,even with a view that the relationships stated aboveare dependent on many parameters such as frequency,load ratio and laminate configuration.

Figure 21: Degradation of matrix-crack fracturetoughness Gmc with increasing number of cycles (bluedots: data points, green line: curve fit).

12

In order to increase the robustness and efficiencyof the proposed fatigue analysis a new approach wasdeveloped and implemented. The main assumptionsof the new formulation will be discussed briefly inthe following. The current damage approach uses theply based material degradation model previously de-veloped by (Shahid and Chang, 1993). This modelconsiders the value of ply matrix-crack fracture tough-ness Gmc as a limit for crack initiation. Tradition-ally, the value of Gmc is assumed being constant un-der static loading conditions. In order to account forthe fatigue loading problem, the matrix-crack fracturetoughness Gmc is introduced here as function of thecycles Gmc = Gmc(N). The value of Gmc willdegrade with increasing number of cycles as shownin Figure 21. The degradation function for the frac-ture toughness Gmc shown by Figure 21 representspreliminary data and has been obtained from back-calculating the matrix-crack fracture toughness fromthe traditional approach results. Finally, a preliminaryempirical relationship was computed by a curve fit ofthis data.Equation 6 shows the mathematical representation ofthe matrix-crack fracture toughness Gmc as a functionof load cycles

Gmc = ANη (6)where N denotes the number of cycles and A, η arematerial constants. This empirical relationship canbe measured using slightly modified standard fracturetoughness testing methods (Nairm, 1999), which arethe subject of future experiments. By introducing Gmcas a function of cycles, the material degradation modelis now dependent on the number of accumulated cyclesN , and accordingly the ply strengths are updated aftereach loading cycle. After each update of the strengthsa modified Hashin strength criterion is evaluated anddamage will propagate if the criterion is fulfilled. Fig-ure 22 illustrates the major steps of the algorithm usedin the current analysis as described.

Figure 22: Algorithm used for the UMAT/Abaqus pro-gressive Finite Element analysis.

In order to validate the proposed algorithm, resultsof the numerical analysis for both the traditional ap-proach using Paris Law and the new approach werecompared with published experimental data (Lafarie-Fernot et al., 2001), (Nairm, 1999). The example pre-sented here is a T300/914 (and T300/934) [0 4/90/03]Slaminate subjected to tension-tension low cycle load-ing. The first step was to perform the static analysisto acquire the static failure load. Once this load wasknown, the cyclic load was set to eighty percent of thestatic failure load.As can be seen from Figure 23, the results of the newapproach clearly give a better representation of the ex-perimental results and follow the experimental trendvery well. Even though there is still room for fur-ther improvement, the current analysis helps us to un-derstand and to characterize damage propagation andstrength degradation for composites. Additionally, itprovides a preliminary tool to develop the DP linkmodules as a major part of the near-future work.

Figure 23: Comparison of numerical and experimentalresults for fatigue analysis; material properties basedon T300/914 (and T300/934).

7 CONCLUSIONSNext generation health management technology willlink different key technologies as impact monitoringby passive sensing, numerical prediction of impact-induced damage, diagnosis of structural damage byactive sensing, and prognostics of remaining usefullifetime to an interactive system for integrated vehiclehealth management. One of the striking features of theproposed system is the continuous exchange of healthdata between the different subsystems to update thestructural models and to feed the real-time automatedreasoning and decision making systems with varioushealth data.The contribution has reported the current state of thedevelopment of an interactive health management sys-tem in terms of system design, basic functionalitiesof the single SHM subsystems and data exchangebetween the modules. Some typical examples havedemonstrated the fundamental capabilities of the pro-posed health management system and open questionsas well as future needs and developments have beendiscussed.

13

In summary, the following capabilities are mergedwithin a unique system for integrated vehicle healthmanagement:

1. Both impact location and impact force history canbe identified in real-time by impact monitoringbased on passive sensing data. An automatedprocedure for system calibration will be imple-mented.

2. The impact characteristics will be submitted sub-sequently to a damage prediction module to es-timate the occurrence and severity of impact-induced damage.

3. After an adverse event, active diagnosis is per-formed to acquire the current health status of thestructure. Environmental effects on the diagno-sis results are compensated by updated baselinedata. For this purpose, the propagation of diag-nostic waves under environmental conditions issimulated and evaluated.

4. Health data from active sensing is passed to theprognostics module where numerical simulationof progressive failure under future load condi-tions is performed. The remaining useful lifetimewill be estimated by consideration of an allowabledamage/degradation limit.

5. A global health management interface controlsthe system and data flow interactively and pro-vides real-time reasoning and decision-makingcapabilities.

ACKNOWLEDGMENTSThe authors would like to acknowledge the NationalAeronautics and Space Administration (NASA) for thesupport of this work under grants NRA-07-IVHM1-07-0061 and NRA-07-IVHM1-07-0064. In particu-lar, we would like to thank Richard W. Ross (NASALaRC) as the program monitor for his support of theproject. Furthermore, the authors are grateful for col-laboration with Honeywell Inc. and Acellent Tech-nologies. In particular, the authors acknowledge Acel-lent’s contribution and assistance on the passive andactive sensing systems.

REFERENCES(Boller et al., 2009) C. Boller, F.-K. Chang, and

Y. Fujino (Eds.). Encyclopedia of structural healthmonitoring. In John Wiley & Sons, 2009.

(Brammer and Percival, 1970) J. A. Brammer andC. M. Percival. Elevated-temperature elastic moduliof 2024 aluminum obtained by a laser-pulse tech-nique. Experimental Mechanics, 10(6):245–250,1970.

(Chang and Beard, 1997) F.-K. Chang and S. Beard.Active damage detection in filament wound com-posite tubes using built-in sensors and actuators.Journal of Intelligent Material, Systems and Struc-tures, 8:891–897, 1997.

(Chang and Ihn, 2004) F.-K. Chang and J.-B. Ihn. De-tection and monitoring of hidden fatigue crackgrowth using a built-in piezoelectric sensor/actuatornetwork: Part ii: validation through riveted jointsand repair patches. In C. T. Sun et al (editors),ASME - AD, pages 621–630, 2004.

(Chang and Qing, 1997) F.-K. Chang and X. L. Qing.Recent advances in structural joints and repairs forcomposite materials. In Edited by L. Tong andC. Soutis, Kluwer Academic Publisher, pages 101–140, 1997.

(Chang and Roh, 1995) F.-K. Chang and Y. Roh. Ef-fect of impact damage on lamb wave propagation inlaminated composite structures. In C. T. Sun et al(editors), ASME - AD, 46:127–138, 1995.

(Chang and Seydel, 2001) F.-K. Chang and R. E. Sey-del. Impact identification of stiffened compositepanels: Part i: system developments. In Smart Ma-terials Structures. Institute of Physics Publishing,2001.

(Chang et al., 2008) F.-K. Chang, I. Mueller, andK. Lonkar. Integrated passive-active interactive di-agnostic technique based on efficient inverse meth-ods for condition monitoring and damage detectionfor an integrated vehicle health management. InAnnual Report to National Aeronautics and SpaceAdministration (NASA), Department of Aeronauticsand Astronautics, Stanford University, 2008.

(Choi, 1990) H. Choi. Damage in graphite/epoxylaminated composites due to low-velocity impact.In Dissertation, Department of Aeronautics and As-tronautics, Stanford University, 1990.

(Doyle, 1987) J. F. Doyle. Experimentally determin-ing the contact force during the transverse impacton an orthptropic plate. Journal of Sound and Vi-bration, 118:441–448, 1987.

(Fogg et al., 1993) B. R. Fogg, M. F. Gunther, andA. Wang. Fiber optic impact detection and loca-tion system embedded in a composite material. InProceedings of SPIE Conference, pages 262–269,1993.

(Geubelle and Baylor, 1988) P. H. Geubelle and J. S.Baylor. Impact-induced de-lamination of compos-ites: a 2d simulation. Composites, Part B, 29B:589–602, 1988.

(Ha et al., 2009) S. Ha, A. Mittal, K. Lonkar, and F.-K. Chang. Adhesive layer effects on temperature-sensitive lamb waves induced by surface-mountedpzt actuators. In Proceedings of 7th InternationalWorkshop on Structural Health Monitoring, Stan-ford CA, pages 2221–2233, 2009.

(Hadamard, 1923) J. Hadamard. Lectures on thecauchy problem in linear partial differential equa-tions. In University Press, New Haven, 1923.

(Harris, 2003) B. Harris. Fatigue in composites. InWoodhead Publishing, 2003.

(Hashin, 1980) Z. Hashin. Failure criteria for unidi-rectional fiber composites. Journal of Applied Me-chanics, 47:329–334, 1980.

(Jiang et al., 1998) Y. P. Jiang, K. Chandrashekara,and A. C. Okafor. Estimation of contact force oncomposite plates using impact induced strain andneural network. Composites Part B: Engineering,29(4):363–370, 1998.

(Kim et al., 2008) Y. Kim, S. Ha, and F.-K. Chang.Time-domain spectral element method for built-in

14

piezoelectric-actuator-induced lamb wave propaga-tion analysis. AIAA Journal, 46:591–600, 2008.

(Komatitsch and Vilotte, 1998) D. Komatitsch andJ.P. Vilotte. The spectral element method: An ef-ficient tool to simulate the seismic response of 2dand 3d geological structures. Bulletin of the Seis-mological Society of America, 88:368–392, 1998.

(Lafarie-Fernot et al., 2001) M. C. Lafarie-Fernot,C. Henaff-Gardin, and D. Gamby. Matrix crack-ing induced by cyclic ply stresses in compositelaminates. Composites Science and Technology,61:2327–2336, 2001.

(Lessard, 1989) L. B. Lessard. Compression failure inlaminated composites containing an open hole. InDissertation, Department of Aeronautics and Astro-nautics, Stanford University, 1989.

(Ljung, 1999) L. Ljung. System identification – the-ory for the user. In PTR Prentice Hall, 2nd edition,1999.

(Markmiller, 2007) J. Markmiller. Quantification andoptimization of a structural health monitoring sys-tem for impact detection in composite structures. InDissertation, Department of Aeronautics and Astro-nautics, Stanford University, 2007.

(Mueller and Chang, 2009) I. Mueller and F.-K.Chang. Model-based impact monitoring by inversemethods using particle swarm optimization. InProceedings of IMAC XXVII, 2009.

(Nairm, 1999) J. A. Nairm. Applications of finitefracture mechanics for predicting fracture events incomposites. In 5th International Conference on De-formation and Fracture of Composites, London UK,1999.

(Park and Chang, 2003) J. Park and F.-K. Chang.Built-in detection of impact damage in multi-yayered thick composite structures. In Proceedingsof 4th International Workshop on Structural HealthMonitoring, Stanford CA, pages 1391–1398, 2003.

(Park, 2005) J. Park. Impact identification in struc-tures using a sensor network: The system identi-fication approach. In Dissertation, Department ofAeronautics and Astronautics, Stanford University,2005.

(Raghavan and Cesnik, 2008) A. Raghavan andC. E.S. Cesnik. Effects of elevated temperature onguided-wave structural health monitoring. Journalof Intelligent Material Systems and Structures,19(12):1383–1398, 2008.

(Scalea and Salamone, 2009) F. Lanza Di Scalea andS. Salamone. Temperature effects in ultrasoniclamb wave structural health monitoring systems.The Journal of the Acoustical Society of America,124(1):161–174, 2009.

(Seydel and Chang, 1999) R. E. Seydel and F.-K.Chang. Implementation of a real-time impact iden-tification technique for stiffened composite panels.In Proceedings of 2nd International Workshop onStructural Health Monitoring, Stanford CA, 1999.

(Shahid and Chang, 1993) I. Shahid and F.-K. Chang.Progressive failure analysis of laminated compos-ites subjected to in-plane and shear loads. In De-partment of Aeronautics and Astronautics, StanfordUniversity, 1993.

(Sherrit et al., 1999) S. Sherrit, G. Yang, H. D. Wied-erick, and B. K. Mukherjee. Temperature depen-dence of the dielectric, elastic and piezoelectric ma-terial constants of lead zirconate titanate ceram-ics. In Proceedings of Intinternation Conferenceon Smart Materials, Structures and Systems, Ban-galore, India, 1999.

(Su et al., 2006) Z. Su, L. Ye, and Y. Lu. Guidedlamb waves for identification of damage in com-posite structures: A review. Journal of Sound andVibration, 295:753–780, 2006.

(Wang and Chang, 2000) C. Wang and F.-K. Chang.Diagnosis of impact damage in composite struc-tures with built-in piezoelectric network. Proceed-ings of Smart Structures and Materials, 3990:13–19, 2000.

(Wu and Chang, 1989) H.-Y. T. Wu and F.-K. Chang.Transient dynamic analysis of laminated compositeplates subjected to transverse impact. Computersand Structures, 31(3):453–466, 1989.

(Wu and Springer, 1988) H.-Y. T. Wu and G. S.Springer. Impact induces stresses, strains and de-laminations in composite plates. Journal of Com-posite Materials, 22:533–560, 1988.

(Yen et al., 1994) C. S. Yen, E. Wu, and J. C. Yeh.Identification of impact forces at multiple locationson laminated plates. AIAA Journal, 32:2433–2439,1994.

15