Real-time impact identification algorithm for composite structures using fiber Bragg grating sensors

Transcription

1 STRUCTURAL CONTROL AND HEALTH MONITORING Struct. Control Health Monit. 2012; 19: Published online 10 April 2012 in Wiley Online Library (wileyonlinelibrary.com) Real-time impact identification algorithm for composite structures using fiber Bragg grating sensors Byeong-Wook Jang 1, Yeon-Gwan Lee 2,Jin-HyukKim 2, Yoon-Young Kim 2 and Chun-Gon Kim 2, *, 1 Korea Aerospace Research Institute, Gwahangno, Yuseong-gu, Daejeon, , South Korea 2 Department of Aerospace Engineering, KAIST, 335 Gwahangno, Yuseong-gu, Daejeon, , South Korea SUMMARY In composite structures, low-velocity impact-induced damage such as delamination are mostly hidden inside laminates or leave a small dent on the impact side. Thus, detecting this type of damage using conventional inspection methods is not easy and requires much time and cost. To enhance efficiency of these methods, information on the estimated impact locations must be provided with a high accuracy. In this way, unnecessary inspections for large intact regions can be reduced. In this study, impact localization algorithms for various composite structures were developed using the impact induced acoustic signals acquired by multiplexed fiber Bragg grating (FBG) sensors. The acoustic waves from a given impact were transmitted to each FBG sensor, and incurred FBG wavelength shifts were captured by a high speed multiplexible FBG interrogation system with a sampling frequency of 100 khz. After acquisition of the FBG sensor signals at all the training points in the target section, the impact wave arrival time differences between each FBG signal were calculated to produce the input data sets for neural network training. To reliably use the neural network algorithm for impact identification, high reproducible arrival time determination algorithms are essentially required. In this study, such arrival time determination algorithms were developed through various types of structures. Finally, we evaluated the performances of the suggested impact identification algorithms for a composite wing box structure. Copyright 2012 John Wiley & Sons, Ltd. Received 10 October 2011; Revised 28 February 2011; Accepted 6 March 2012 KEY WORDS: composite structures; impact identification; high speed FBG strain sensors; arrival time determination; neural network training 1. INTRODUCTION The use of composite materials as primary constituents in commercial, civil, and military applications has steadily been increasing. Above all, the use of composites in aerospace industries has attracted much attention because of their ability to reduce structural weight. Consistent with this trend, concepts of structural health monitoring (SHM) have been studied to ensure reliability and safety, and to efficiently inspect and maintain the composite structures [1]. SHM techniques are expected to become more important in the change from conventional schedule-based inspections to condition-based inspections. Furthermore, realtime damage detection during operations is also critical in SHM techniques. At this time, SHM techniques are being gradually applied to military aircrafts, particularly composite unmanned aerial vehicles (UAVs) [2]. Recently, health and event monitoring in UAVs have become important considerations as structural design aspects and criteria [3]. Moreover, decision making systems on continuing specified missions are required in UAVs because there are no pilots to monitor events and structural status. For these reasons, *Correspondence to: Chun-Gon Kim, Department of Aerospace Engineering, KAIST, 335 Gwahangno, Yuseong-gu, Daejeon, , South Korea. cgkim@kaist.ac.kr Copyright 2012 John Wiley & Sons, Ltd.

2 REAL-TIME IMPACT IDENTIFICATION ALGORITHM FOR COMPOSITE STRUCTURES 581 SHM techniques can provide efficient solutions as built-in non-destructive testing (NDT) systems in UAV structures. Among the various types of damage that need to be monitored on targets, low-velocity impact-induced damage such as delamination or matrix cracks, are mostly hidden inside the laminates or occur in the opposite side of the impact. Thus, without information about the impact locations, all suspicious regions must be inspected, and this results in inefficiencies in the inspection process. To reduce the time and costs that result from unnecessary inspections, information about impact locations must be estimated appropriately and provided to the related NDT operators. Since 1990, there has been much research surrounding impact location detection for various types of structures. Gaul et al. [4] used four PZT (lead zirconate titanate) sensors to detect impact locations on a steel plate by measuring the impact wave speed. Seydel et al. [5,6] undertook impact identifications for graphite composite panels with stiffened ribs using the Stanford multi-actuator-receiver transduction layer technique. Sung et al. [7] used the neural network algorithm to estimate impact locations on a graphite composite plate. Sekine et al. [8] suggested multiple-impact identification methods for composite isogrid-stiffened panels using a genetic algorithm and numerical methods. In addition, Ciampa et al. [9] and Hajzargerbashi et al. [10] suggested the advanced triangulation methods for impact localizations with reducing the nonlinear effects of impact wave speed in anisotropic materials. Chen et al. [11] adopted a time-reversal method to identify the impact source locations using a finite element analysis. Wang et al. [12] tried to develop a remote identification of impact damage, and other researchers [13,14] used fiber Bragg grating sensors for monitoring of impact damage. Although most efforts produced good estimated results, there are some limitations in applying these techniques to real structures because of the complexity of sensing and signal processing systems, electromagnetic interference problems associated with electrical sensors, relatively low sampling frequencies [15] and multiplexing problems of optical fiber sensors [16]. Thus, numerous researches [17,18] about impact and damage localizations are still in progress. In this study, to simplify the sensing components, multiplexed FBG sensors were applied by surface attachments on the target structure. Although much effort [19,20] has been tried to obtain high speed FBG interrogations with multiplexing, there are still limitations because of a low sampling frequency (under 50 khz). The limitation of a low sampling frequency was overcome using the newly developed FBG interrogation system based on spectral domain measurements [21]. To increase the applicability of the impact identification methods, the impact experiments in this study were performed on various types of composite structures such as a composite flat plate, a composite stiffened panel [22] and a full scale UAV composite wing box [23]. In this case, the low signal-to-noise ratios of the FBG sensor signals disturb to obtain the reproducible arrival times as the neural network input data. To reduce this undesirable effect, various arrival time determination algorithms were suggested. Then, from the neural network training using the obtained arrival time data on the training grid points, trained weight factors were obtained. Finally, these impact location estimation methods were verified using validation tests on the non-trained grid points. To explain the procedures of developing the suggested impact identification algorithm, the research results in our previous work [22] are included in this paper. 2. LOW VELOCITY IMPACT EXPERIMENTS 2.1. High speed fiber Bragg grating interrogation system Fiber Bragg grating sensors have the advantage of using a distributed sensing scheme because of their multiplexing capability. However, the direct measurement method of the Bragg wavelength is limited in obtaining high frequency signals such as impact-induced acoustic emissions. In addition, the demodulation methods that use a tunable laser or Fabry Perot filter [24,25] have the disadvantages of restricting the detectable frequency range because of the number of multiplexed sensors, system complexity, and high cost. Thus, as essential requirements of an efficient and simple impact measurement system based on FBG sensors, a sufficient sampling rate and a multiplexing capability over four sensing points must be included together. Recently, FIBERPRO Inc. (Korea) developed a multiplexible FBG wavelength interrogator with a sampling rate of 100 khz. This interrogation system consists of a superluminescent laser diode (SLD), a spectrometer, a photodiode array, and fully parallel readout circuits. When the lights reflected from the FBG sensors are dispersed by a bulk phase grating, the photodiode array converts the imaged optical

3 582 B.-W. JANG ET AL. signals to electrical signals. These electrical signals are quickly processed using parallel photocurrent amplifiers and analog-to-digital converters following each photodiode (80 photodiodes in the spectral domain from 1530 to 1560 nm). Through the above processing principle and centroid method for finding the central wavelength, this interrogation system can measure eight FBG signals simultaneously with a sampling frequency of 100 khz. Among the currently available technologies related to high speed FBG interrogators, this system demonstrates better performances in terms of multiplexing, sampling rate, and system simplicity. The specifications are summarized in Table I. In the low velocity impact experiments in this study, the high speed FBG interrogation system detailed above was used to acquire the impact signals from the multiplexed FBG sensors. We expect that this interrogator will contribute to the construction of more applicable impact monitoring systems Specimens and experimental setup To develop the impact identification system for composite structures, we carried out low velocity impact experiments in three types of structures. As a first step, impact experiments were performed on specimen levels, namely a composite flat plate and a composite stiffened panel as depicted in Figures 1 and 2 [22]. In both cases, four multiplexed FBG sensors (four sensor heads in one single fiber) were used, and each sensor was attached to the bottom surface of each corner in the test section. Only sensor regions of a whole optical fiber were bonded using cyanoacrylate adhesive with surface treatments. From these experiments, preliminary impact identification algorithms would be suggested. In this step, impact signals were sampled at 40 khz because the high speed interrogator was still being developed at that time. Table I. Specifications of the high-speed fiber Bragg grating interrogation system. Specification Value Wavelength range 1530 ~ 1560 nm (30 nm) Wavelength accuracy <30 pm Wavelength repeatability <10 pm Dynamic range >10 db Number of detectable sensors Up to 6 (5 nm spacing) Number of channels 1 Max. measurement frequency 100 khz Optical connector type FC/APC Optical source SLD Operating temperature 5~50 C Humidity 20 ~ 85 % RH FBG 2 FBG m FBG 3 FBG m 0.69 m Figure 1. Composite flat plate specimen and locations of fiber Bragg grating sensors.

4 REAL-TIME IMPACT IDENTIFICATION ALGORITHM FOR COMPOSITE STRUCTURES 583 FBG 4 FBG 3 y FBG 1 FBG 2 (0,0) x 0.1 m Figure 2. Composite stiffened panel specimen and locations of fiber Bragg grating sensors. For modifications of such impact identification algorithms for real applications, experiments on a real composite structure were considered. The full scale UAV composite wing box structure and test section are shown in Figure 3. It is composed of upper and lower skins, three spars (front, intermediate, and rear), ribs, and aluminum fitting lugs. The composite wing section was designed and manufactured by DACC Ltd. (Korea) for the SHM project titled Korean aero-vehicle structural health monitoring system (KASHMOS). The KASHMOS project is being led by Agency for Defense Development Test section FBG 6 FBG 1 FBG 5 FBG 2 2,340 mm 1,117 mm FBG 4 FBG FBG 1 FBG 6 Y (m) FBG 2 FBG FBG 3 FBG X (m) Figure 3. Configuration of the full-scale unmanned aerial vehicle composite wing box structure.

5 584 B.-W. JANG ET AL. (ADD, Korea) and focuses on developing onboard and ground SHM systems for composite UAVs. This study is included in the scope of the KASHMOS project. The test section was m 2 on the upper skin of a wing box, and the grid size was m 2. This upper skin was fabricated using a graphite/epoxy prepreg (USN 175BX, SK Chemicals, Korea) and the stacking sequence was [45/0/ 45/90/ 45] s. Six multiplexed FBG sensors (six sensors in one single fiber) were attached to the surface of the bottom side, but only four sensors at each corner side were used for impact localizations. In order to produce low-velocity impacts without dual impacts, an instrumented impact test fixture was used. The impact angles were adjusted to be perpendicular to the test article by tilting the fixture. Furthermore, the impacts were made by free drop, and the energy was controlled by controlling the drop height. In these experiments, an impact energy of 2 J was provided to prevent impact damage to occur. In this experiment, the FBG signals were obtained at a sampling frequency of 100 khz using the high-speed interrogator Fiber Bragg grating impact signals Figure 4 shows the signals acquired by the FBG sensors from the impact on the point close to FBG 1 in each structure. These figures confirm that impact events were appropriately captured by the high speed FBG interrogator. According to the distance between the sensor and the impact location, the amplitudes of the wavelength shift in FBG 1 are larger than those of the other sensors. In addition, the leading waves arrive earlier at FBG 1, and the arrival times are similar in the equidistant sensors. The confirmations of the other point data show that the arrival times of each leading wave were sufficiently distinguished because of the distances between the impacts and sensor locations. Thus, we concluded that the FBG sensor signals acquired from the interrogation system are useful for detecting impact locations in this research. Amplitude (nm) FBG 4 FBG FBG FBG Time (sec) Time (sec) (a) Amplitude (nm) FBG 4 FBG 3 FBG 2 FBG 1 (b) Amplitude (nm) 0.10 FBG 6 FBG FBG 3 FBG Time (sec) (c) Figure 4. Obtained fiber Bragg grating signals: (a) composite flat plate, (b) composite stiffened panel and (c) composite wing box.

6 REAL-TIME IMPACT IDENTIFICATION ALGORITHM FOR COMPOSITE STRUCTURES PRELIMINARY IMPACT IDENTIFICATIONS 3.1. Arrival time determination algorithm For using the arrival time differences as the input data for neural network training, a reliable arrival time determination algorithm is the most important requirement. Before working with a real structure, we first investigated the arrival time determination method in specimen levels. In anisotropic materials such as composites, there are several propagation modes with different frequency bands in an impact wave. Thus, for proper determinations of arrival time differences, arrival times had to be measured from the leading waves or similar frequency portions of each sensor signal. In order to choose the specific propagation mode, which appropriately reflects the arrival moment of an impact wave, a wavelet transform was applied to the obtained signals. In this study, the Daubechies 4 (db 4) wavelet functions were adopted to decompose the sensor signal into four wavelet portions (D 1,D 2,D 3, and D 4 ). Among the decomposed wavelet portions, the D 2 portion properly indicates the arrival time of the original signal as shown in Figure 5. Moreover, the noise level of D 2 portion is lower than that of an original signal, so that the arrival moment of an impact wave could be more easily and reliably distinguished Original signal D2 wavelet decomposed portion Wavelength shift (nm) Time (sec) Figure 5. Comparison between the original signal and the wavelet decomposed portion. Data acquisitions Baseline converting Wavelet transform Subtraction of D 2 detailed portion Arrival times Threshold level Figure 6. Flow chart for the preliminary arrival time determination algorithm.

7 586 B.-W. JANG ET AL. Figure 6 shows the flow chart for the suggested arrival time determination algorithm. In the first step, the baselines of the raw signals were converted taking the central wavelength values into account before the impact events. Then, a wavelet transform was applied to the baseline converted signals to obtain D 2 portions as similar frequency modes. Using such D 2 portions, each arrival time was determined when the data value was over one and a half times the noise level of a starting signal before an impact event (t i : arrival time of FBG 1). Using this algorithm, the arrival time differences (dt 1 =t 2 t 1,dt 2 =t 3 t 2, and dt 3 =t 4 t 3 ) were obtained from all the training grid points. The results of repeated tests confirmed that this process offers reasonable arrival time and shows good reliability Impact identifications using neural network training The neural network used in this study adopted a multi-layered perceptron which consisted of input, hidden layers, and output layer. There were three, six, and two nodes in each layer, respectively. The steepest descent method was used as the back-propagation learning algorithm. The input data were the arrival time differences, and the output data were the impact location as shown in Figure 7. Each pair of input and output data formed a pattern, and the number of total patterns was the same as the number of training points for each structure. There were three data sets because the impact experiments were performed three times. Among these data sets, data set 1 was trained to acquire the weights and bias matrices. Through the back-propagation learning procedure, the pairs of input and output data are repeatedly trained until the error between calculated output data and expected values is under the criterion. As a datum value for the error, the root mean squared error (RMSE) is adopted. For calculating the RMSE value, differences between the output and expected values are firstly squared, and then such squared values are summed and divided by the number of training points. In every epoch (whole patterns) level learning, the RMSE value is calculated and compared with the criterion for determining the stop of training. Until the calculated error is under the criterion, all weight factors (w ij, b j, w jk,andb k ) are updated using the RMSE value. When the expected output values are obtained, the training procedure is terminated, and the trained weight factors are obtained. Using such trained weight factors and arrival time difference data from the obtained sensor signals, the impact location can be estimated. Figure 8 shows the detected impact locations using the trained data set for the composite flat plate and composite stiffened panel. As a result, the maximum error was mm, and the average error was 8.20 mm for the composite flat plate. In the composite stiffened panel, the maximum error was Hidden Layer (1) y j Input Layer (1) w ij (2) w jk Output Layer t 2 t 3 x y t 4 (1) b j (1) b k Bias 1 Bias 2 Figure 7. Multi-layered perceptron neural network.

8 REAL-TIME IMPACT IDENTIFICATION ALGORITHM FOR COMPOSITE STRUCTURES Real locations Real locations Detected Detected locations locations 0.5 Real locations Detected locations Real locations Detected locations Y(m) 0.10 Y(m) X(m) (a) X(m) Figure 8. Results of impact identifications in the composite flat and stiffened specimens: (a) composite flat plate and (b) composite stiffened panel. (b) mm, and the average error was mm. The detected impact locations deviated less than each grid size, so that the preliminary impact identification algorithm functioned well Verifications In order to verify the proposed algorithm, the impact locations on arbitrary points in the test section must be detected. We performed the verification tests for the composite stiffened panel. Among the non-grid points, four points identified in Figure 9 were determined as the validation points. From the repeated impact experiments, the FBG signals were acquired and the arrival times were determined. The results of the impact identifications of these validation points are shown in Figure 9. The maximum and average errors of the detected results were mm and mm, respectively. The calculated errors were smaller than the grid size (0.1 m). Thus, each estimated impact location was sufficiently distinguished from the other points. Although these results were not much better than those of previous research, it is meaningful because this study used a simple fiber optic sensor system and operated in more complex and larger structures. 0.5 Real locations Detected locations 0.4 Y(m) Point 1 Point Point 2 Point X(m) Figure 9. Verification results of impact identifications in the stiffened panel specimen.

9 588 B.-W. JANG ET AL. 4. IMPROVEMENTS AND APPLICATION 4.1. Improvements of preliminary impact identification algorithm For an efficient impact monitoring, a higher sampling frequency of over 50 khz is required because the sampled data over 50 khz can be used for damage assessments as well as impact identifications. For this reason, FBG data were sampled at 100 khz using an upgraded FBG interrogation system for a composite wing box. Before applying the preliminary arrival time determination algorithm, changes in signal characteristics attributable to different sampling frequencies (40 and 100 khz) had to be investigated. First of all, the noise level of the acquired signals is the most significant factor because such signals can be the main cause of unexpected noise peaks that can induce uncorrected arrival time detection. In order to check the noise levels, sensor data were acquired at different sampling frequencies as shown in Figure 10. As a result, the magnitudes of peak-to-peak values before arrival of the impact wave at 100 khz were larger than those in 40 khz. This noise level also affects the signal-to-noise ratios and creates ambiguous determinations of arrival times. When the preliminary arrival time determination algorithm using the wavelet portion D 2 was applied to the 100 khz sampled signals, fault arrival times were obviously detected. It means that the preliminary algorithm could Wavelength shift (nm) khz 40 khz Time (sec) Figure 10. Comparison between signals obtained at different sampling frequencies. Data acquisitions (n = 4,000) Baseline converting Base signals (i = 0 ~ 50) RMS calculations ii = 0 Subtracted signals ( j = ii ~ ii+50) Arrival times Threshold level ii = ii + 1 ii < n Figure 11. Flow chart for the improved arrival time determination algorithm.

10 REAL-TIME IMPACT IDENTIFICATION ALGORITHM FOR COMPOSITE STRUCTURES 589 not guarantee the repeatable results as well as correct determinations of arrival times. Thus, the arrival time determination algorithm had to be improved for applications to high sampled sensor data. In the improved algorithm (Figure 11), 50 points were subtracted from the baseline-converted signals as base signals before arrival of the impact wave. Then, the root mean squared (RMS) values between the base and baseline converted signals were calculated according to the time axis, so that the RMS curves could be obtained in each FBG sensor signal. As shown in Figure 12, the noise of the RMS curve was reduced to almost half of the original signal. Thus, the arrival times of leading waves could be reliably determined without an additional data process, such as a wavelet transform in the preliminary algorithm. This point is advantageous in realizing the proposed algorithm in programmable circuits for on-board equipment without a computer because the wavelet transform is so complex. Furthermore, the back-propagation learning algorithm in a neural network is changed from the steepest descent to the Levenberg Marquardt algorithm, which is known as one of the fastest learning algorithms. As the number of patterns for neural network training increases, the convergence of errors becomes difficult and requires much computing time. When the steepest descent algorithm was used in the training of a composite wing box case, the output values could not be converged to the expected values over millions of epoch. In order to solve this convergence problem, the Levenberg Marquardt algorithm was adopted. The equation for updating the weights in 0.04 Wavelength shift (nm) Baseline converted signal RMS values Arrival time of FBG Time (sec) Figure 12. Comparison between the original signal and the root mean squared curve. 0.5 Real locations Trained results (set 1) 0.4 Y (m) X (m) Figure 13. Results of impact identifications in the composite wing box.

11 590 B.-W. JANG ET AL Real locations Test results_point 1 Test results_point 2 Test results_point Point Point 2 Y (m) 0.2 Point X (m) Figure 14. Verification results of impact identifications in the composite wing box. the Levenberg Marquardt algorithm is given as the following equation. w kþ1 ¼ w k J T 1J J þ mi T e In this equation, e is the error vector with respect to the weights and biases, and J is the Jacobian matrix composed of the first derivatives of e. The learning rate, m is updated from the changes in the sum of the squared errors. According to the learning rate, a back-propagation algorithm is co-adjusted between the Newton method and steepest descent method. Using this algorithm significantly reduces the convergence time within 100 epochs training Impact identifications in a composite wing box Using the improved algorithm, the impact identifications were performed for a full-scale composite wing box structure as shown in Figure 13. The maximum and average errors of the detected results from all three data sets were mm and mm, respectively. Figure 14 shows the verification results about three non-trained grid point regions. The maximum error was mm, and the average error was mm. These results confirm that the proposed impact identification algorithm appropriately functioned in a real application. 5. CONCLUSION In this research, impact identification algorithms were suggested using the impact-induced acoustic waves acquired by multiplexed FBG sensors in a single optical fiber line. To simultaneously obtain the impact signals from multiple FBG sensors, a newly developed high-speed FBG interrogation system was used. Then, in order to use a neural network for impact identifications, arrival time determination algorithms were investigated to obtain reliable input data from various types of composite structures. As a result, the reproducibility of the arrival time determination results was enhanced, and the impact locations were predicted well even on a very complex structure, a composite wing box. Finally, this suggested impact identification algorithm was verified from the results in the non-trained grid points. In conclusion, this study shows the impact identifications on a full-scaled composite wing box structure. As a result of its simple sensing system and calculation process, our impact identification system appears to be more efficient for use in real structures than systems proposed in previous studies. From these results, we can expect that the suggested impact monitoring system with this impact identification method can be useful for maintaining and repairing composite aerial vehicle structures.

12 REAL-TIME IMPACT IDENTIFICATION ALGORITHM FOR COMPOSITE STRUCTURES 591 ACKNOWLEDGEMENTS The present work was supported by the Agency for Defense Development in Korea (contract no UC080019JD), a grant (07 02) from Aviation Safety R&D Program funded by Ministry of Land, Transport and Maritime Affairs of Korean government and Nuclear Research and Development Program (Grant No ) through the National Research Foundation of Korea funded by the Ministry of Education, Science and Technology. This support is gratefully acknowledged. REFERENCES 1. Measures RM. Structural Monitoring with Fiber Optic Technology. USA: ACADEMIC PRESS, 2001; Park CY, Cho CM, Jun SM. Structural damage monitoring of a composite wing using multiple types of sensors. Advanced Materials Research 2010; : Neubauer M, Gunther G, Fullhas K. Structural design aspects and criteria for military UAV. RTO-MP-AVT 145 UAV Design Processes and Criteria Gaul L, Hurlebaus S. Identification of the impact location on a plate using wavelets. Mechanical systems and signal processing 1998; 12: Seydel R, Chang FK. Impact identification of stiffened composite panels: I. System development. Smart Materials and Structures 2001; 10: Seydel R, Chang FK. Impact identification of stiffened composite panels: II. Implementation studies. Smart Materials and Structures 2001; 10: Sung DU, Oh JH, Kim CG, Hong CS. Impact monitoring of smart composite laminates using neural network and wavelet analysis. Journal of Intelligent Material Systems and Structures 2000; 11: Sekine H, Atobe S. Identification of locations and force histories of multiple point impacts on composite isogrid-stiffened panels. Composite Structures 2009; 89: Ciampa F, Meo M. A new algorithm for acoustic emission localization and flexural group velocity determination in anisotropic structures. Composites: Part A 2010; 41: Hajzargerbashi T, Kundu T, Bland S. An improved algorithm for detecting point of impact in anisotropic inhomogeneous plates. Ultrasonics 2011; 51(3): Chen C and Yuan FG. Impact source identification in finite isotropic plates using a time-reversal method: theoretical study. Smart Materials and Structures 2010; 19: (11p). 12. Wang B, Takatsubo J, Akimune Y, Tsuda H. Development of a remote impact damage identification system. Structural Control and Health Monitoring 2005; 12(3 4): Kirikera GR, Balogun O, Krishnaswamy S. Adaptive fiber Bragg grating sensor network for structural health monitoring: application to impact monitoring. Structural Health Monitoring 2011; 10(1): Hackney D, Peters K. Damage identification after impact in sandwich composites through embedded fiber Bragg sensors. Journal of Intelligent Material Systems and Structures 2011; 22(12): Jensen AE, Havsgard GB, Pran K, Wang G, Vohra ST, Davis MA, Dandridge A. Wet deck slamming experiments with an FRP sandwich panel using a network of 16 fiber optic Bragg grating strain sensors. Composites: Part B 2000; 31: Staszewski WJ, Read I, Foote PD. Damage detection in composite materials using optical fibres recent advances in signal processing. Proc SPIE 2000; 3985: Ciampa F, Meo M. Impact detection in anisotropic materials using a time reversal approach. Structural Health Monitoring 2012; 11(1): Laflamme S, Kollosche M, Connor JJ, Kofod G. Soft capacitive sensor for structural health monitoring of large-scale systems. Structural Control and Health Monitoring 2012; 19: Hongo A, Kojima S, Komatsuzaki S. Applications of fiber Bragg grating sensors and high-speed interrogation techniques. Structural Control and Health Monitoring 2005; 12(3 4): Fujisue T, Nakamura K, Ueha S. Demodulation of acoustic signals in fiber Bragg grating ultrasonic sensors using arrayed waveguide gratings. Japanese Journal of Applied Physics 2006; 45(5B): Lee BW, Seo MS, Oh HG, Park CY. High-speed wavelength interrogator of fiber Bragg gratings for capturing impulsive strain waveforms. Advanced Materials Research 2010; : Jang BW, Park SO, Lee YG, Kim CG, Park CY, Lee BW. Impact monitoring of composite structures using fiber Bragg grating sensors. Journal of the Korean Society for Composite Materials 2011; 24: Choi M, Kim J, Kim Y, Kim J, Choi K. Development of sensor integrated composite wing structure. Advanced Materials Research 2010; : Betz DC, Thursby G, Culshaw B, Staszewski WJ. Acousto-ultrasonic sensing using fiber Bragg gratings. Smart Materials and Structures 2003; 12: Kersey AD, Berkoff TA, Morey WW. Multiplexed fiber Bragg grating strain sensor system with a Fabry Perot wavelength filter. Optics Letters 1993; 18: