Composite Structures 183 (2018) 338–346

Contents lists available at ScienceDirect

Composite Structures
journal homepage: www.elsevier.com/locate/compstruct

Modal acoustic emission for composite structures health monitoring:

Issues to save computing time and algorithmic implementation
S. Yaacoubi ⇑, F. Dahmene, A. Bouzenad, M. El Mountassir, M. Aouini
Institut de Soudure, Plateforme RDI CND, 4 Boulevard Henri Becquerel, 57957 Yutz, France

a r t i c l e i n f o a b s t r a c t

Article history: This paper deals with assessing the integrity of composite structures via modal acoustic emission (MAE)
Received 31 January 2017 technique. It is a continuity of the paper ‘‘On the modal acoustic emission of composite structures”, pub-
Accepted 22 March 2017 lished last year by the authors in the same journal. To improve the reliability level of this technique (as it
Available online 29 March 2017
is the case of the other nondestructive techniques), exploring various innovative processing techniques of
the collected data is required. Unfortunately, this task undergoes a huge volume of data, where its man-
Keywords: agement (processing, reuse, etc.) should be achieved as well as possible. Hence, performing an efficient
Composite structures
processing of the large data sets that can be generated via MAE, during the composite structures health
Modal acoustic emission
Monitoring technique
monitoring, is a challenging topic. This study concerns the development of an algorithmic tool for resolv-
Algorithms ing the problem of memory saturation, which can be encountered when working with such large data
Computing time sets. It is a first step towards Big Data based solutions, launched recently by the authors’ team. A case
Memory mapping of study is discussed, showing the robustness of the already implemented algorithmic tool.
Ó 2017 Elsevier Ltd. All rights reserved.

1. Introduction ing plies and at a certain level of stress, a new crack appears. Some
experimental and numerical works have been established the
Composite structures are in permanent increase use in many influence of cracks density on stiffness changes [10]. Usually, stiff-
fields such as vessels in transport and energy [1,2], large shell in ness change due to TMC is not critical as such but it precedes other
aeronautics [3], tubes in oil and gas [4], bars in civil infrastructures critical damage mechanisms [11], and its detection by means of
[5], etc. Mechanical performances of this kind of structures depend non-destructive techniques could be used as an early warning to
on various parameters related to the loading pattern, structural avoid catastrophic failures [12]. During the charge transfer from
design (winding angles, lay-ups sequence. . .) and also to the dam- the cracked layers to the adjacent ones, delamination may be cre-
age state. Damage mechanisms are manifold but the main types ated when loading exceed the yield strength or delamination resis-
are generally matrix cracking, delamination between layers and tance. Fiber-bundle breakage is the most severe kind of damage
fiber breakage. It is well known that damage develops early in ser- that can be dangerous for the structure [13]. Fiber strength is very
vice and accumulates during service life [6]. Depending on the load high and its breakage is the result of a high loading pattern along
and/or mechanical impact characteristics, damage type and rate the fiber axis [14].
will be different and could have an influence on the ultimate fail- Identification of damage mechanisms using a non-destructive
ure [7]. For this reason, the study of the tolerance to damage and technique is essential in the life cycle of such a composite struc-
the relationship between mechanisms and the reduction of stiff- ture. This helps ensuring a more efficient preventive maintenance
ness is a crucial task. For example, when laminates composite is since it is to be achieved in an early stage of eventual defects. This
subjected to flexural fatigue loading, transverse matrix cracks offers a pragmatic benefit to the maintenance team, which is hav-
(TMC) is generally the first damage that occurs in the outer plies ing enough time to react. The more realistic decision to be made is
and grows through the thickness to the other group of plies [1,8]. achieving more regularly this non-destructive testing. Basing on
The multi-cracking is another damage mechanism, which is well the aforementioned damage mechanisms analysis, increasing the
described in literature such as in Garret and Bailly [9]. When a frequency of site intervention (to do measurements) allows deter-
matrix crack appears, a part of the load is transferred to neighbor- mining the velocity of damage evolution. This would be a primor-
dial input to predict the structure residual life. We speak hence
about predictive maintenance, which is a relatively new thematic
⇑ Corresponding author.
that starts gaining ground the last years in many fields. Moreover,
E-mail address: [email protected] (S. Yaacoubi).

http://dx.doi.org/10.1016/j.compstruct.2017.03.081
0263-8223/Ó 2017 Elsevier Ltd. All rights reserved.
 S. Yaacoubi et al. / Composite Structures 183 (2018) 338–346 339

another main advantage is that knowing damage mechanisms (and agation. For this raison, one event generating one echo, at a given
so the root-cause of failures) should be helpful to improve the time, may give rise to a signal consisting of several echoes. For
design of the next future generation and possibly better managing illustration, let us examine Fig. 1 (left), which presents a typical
the load solicitations, to be more robust and so, more safe for assets signal. As it can be seen, two echoes (i.e. wave packet) are present;
and people. Modal acoustic emission (MAE) [15] is a potential can- this means that the event which generated this signal (said also
didate for reaching these aims. The main advantage of this tech- waveform) provided two different modes. While AE uses this
nique is that it offers possibility to perform a global testing or specificity simply as a number of counts (in the current example,
monitoring of a given structure using sparse sensors, and so win this number is equal to two), MAE aims to exploit it (i.e. wave
time and cost (comparatively to the other local techniques such modes types) for determining source orientation and thus, identi-
as ultrasonic phased array [16,17]). fying damage mechanisms. The relationship between source orien-
Unfortunately, although it is a promising technique as it is tation and the generated mode type is largely studied in literature
shown in [15], its implementation is not a trivial item. The ultimate [6].
objective of this paper is so to detail the algorithmic aspects that Theoretically speaking, there is infinity of guided wave modes,
permit achieving MAE. It should be noted that when operating see textbooks of mechanical guided waves such as [20]. In practice
with MAE as well as classic AE, data analyst are confronted to a however, only few modes can propagate and be received by the
huge amounts of data. The challenge is then how to overcome this used sensor. This is due to many factors: the type of the damage
problem by automatizing as match as possible the analysis of the mechanism, the attenuation of waves in the medium, the type of
collected data, and so having a quick access to the result. All this the sensor and its frequency bandwidth, etc. In MAE, the modes
should be executed via a relatively classic microcomputer. The to be exploited are fewer and are namely the symmetric mode
problem that the data analyst should face in the current case is, (called also extensional or compressional) and the asymmetric
as it were, a problem of Big Data [18], but at a relatively small- (named also flexural). Compressional and flexural modes are in
scale. reality two families or modes, where each one consists of many
This paper consists of six sections. The second one is devoted to modes with different orders. In MAE, only the predominant modes
better explain the concept of MAE technique as well as the motiva- are those used for analyses. The lowest symmetric order mode may
tion of achieving this work and writing this paper. Section three is be generated by matrix cracking and fiber fracture whereas the
dedicated to describe the experimental set-up and approach used lowest asymmetric order mode is generated by delamination. By
in this work. Section four is reserved to detail the algorithmic analysing the energy and frequency content of waveforms, it is
implementation. Results are discussed in section five. Conclusions possible to sort the predominant mode and thus identify the
and outlooks of this study are the target of the last section. source. Many works can be found in literature [21] where the more
recent (to the knowledge of the authors) is the one published one
2. Background and motivation year ago by them [15], in which a novel method to assess damage
mechanisms has been proposed.
This section concerns the definition of the MAE and the motiva- The main technical problem that should be encountered when
tion of running this study. The fact that MAE is a relatively recent appealing the proposed method is its numerical implementation.
technique rends the difference between it and the well-known AE Indeed, a composite structure is, by definition, a very high emissive
[19] ambiguous. Actually, the difference is slight and is somewhat medium. During a relatively short time of monitoring (whatever
based on some subtleties. MAE can be considered as a derivative of with MAE or AE), many millions of signals can be recorded. Many
AE, the mother technique. From the point of view of physics of Giga octets can be recorded in just one experiment and one small
waves’ propagation, both techniques are based on guided waves. structure. Fig. 1 (right) is given here for a purpose of illustration,
When generated by an event (such as a damage mechanism), where a very high density of points can be seen. Every point corre-
waves will propagate in a structure with possible various modes sponds to the maximum amplitude of one signal. That is to say that
of propagation [20]. These modes have different velocities of prop- each data array (i.e. signal) is transformed in only one datum. This

Fig. 1. Schematic representation of a typical acoustic emission waveform (left) and acoustic emission activity (right) (CHAN indicates channel; 4 channels were used to
perform acquisitions).
340 S. Yaacoubi et al. / Composite Structures 183 (2018) 338–346

practice is often used in operational monitoring of composite After cut, the obtained specimens were tabbed at the both ends
structure using AE, and that to avoid recording waveforms to face with High density Polyethylene (PEHD) tabs using high strength
the problem of voluminous data. Mega data causes many various adhesive Loctite 3038 (Fig. 2 right). Tabs were used in order to
problems: safeguard, back-up, computing, etc. Fig. 1 right present avoid damage formation under the grips.
an example of a huge amount of data and argues the necessity to
deal with. Additionally, the method proposed regenerate data 3.2. Modes separation
because each signal is transformed in an image (i.e. one data array
becomes a data matrix). If a signal contains N samples, the image Before tensile tests with AE recording, pencil leads were broken
could contain N2 samples. If so the data to be analysed consists on both the edge and surface of a typical coupon. Two wide band-
of millions of signals, it should be developed on N⁄millions of sam- width (WD) sensors are positioned at 100 mm from the top of the
ples during computing! In another words, a one Giga octet data specimen, at the same point but on opposite sides (Fig. 4). These
may become 100 Giga octets! It is so impossible to run such calcu- measurements were used to investigate the frequency content of
lations with a personal computer. This should require a specific extensional and flexural modes. Frequency range of both modes
sophisticated computing station, which might be costly and this was determined using Wavelet Transform. Since flexural mode is
don’t help so democratisation of the MAE. The present paper characterized by a low frequency range [25–55] kHz and the
addresses this problem and presents the algorithmic technique extensional mode, by a high frequency range [125–300] kHz,
that allows implementing the proposed method through a micro- extraction of the predominant mode can be performed using fre-
pc. quency analysis. As the predominant mode is directly linked to
It should be noted moreover that the paper already published the source orientation, modal analysis of waveforms may be used
by the authors [15] and the current one can be considered as the to determine the damage mechanism in composite. This work
first steps of building an accurate MAE technique procedure for was detailed in a previous paper [15].
determining the damage mechanisms in composite vessels. It
should be noted however that the methodology is not limited only 3.3. Tensile test with AE
to vessels. It is applicable to any kind of composites. Indeed, the
collected signals will certainly change of shape but the concept The tensile tests were carried out using a pre-delaminated spec-
of wave modes will remain since this one is property of the geom- imen in order to identify damage mechanisms by MAE and in par-
etry of the waveguide (i.e. the structure) and not its material. ticular to sense the delamination growth. The tests were carried
out using a Zwick machine with a full-scale load range of 250 kN.
3. Experimental approach and set-up The load-displacement data was recorded at a constant crosshead
speed of 1 mm/min.
The experimental methodology of this research work has Various types of AE sensors were strategically placed over the
involved various stages such as samples preparation, modes sepa- specimen in specific locations in order to sense the principal
ration, and tensile test using acoustic emission. These main three delamination growth and eliminate all grid and external noise
steps are detailed hereafter. coming from outside the specimen (Fig. 5). To do that, two reso-
nant sensors were placed from either side of the specimens acting
3.1. Specimens preparation as a guard sensors (all activity detected by these sensors is elimi-
nated). Two WD sensors were placed symmetrically to the thick-
The specimens on which experiments were carried out are par- ness specimen, near the delamination. Details of the hardware
allelpepidic, having 10 mm
36 mm
250 mm as dimensions. and software and sensors for the AE recording are in Table 1. In
These dimensions were established according to ISO527-4 stan- order to investigate only waveforms coming from the region of
dard. Theses specimens were extracted from a relatively big vol- interest (closely to the delamination), a zonal approach was used.
ume composite vessel, which was especially dedicated to To do this, only waveforms received first by the sensor n°3 and
hydrogen storage and transport. The said composite vessel was then by sensor n°2 were considered. All waveforms detected by
manufactured by Stelia Composites Company. It is made of an these sensors are recorded and analysed by means of the algorithm
inner plastic liner and two outer composite overwrapping layers: detailed in the following section (see Figs. 6 and 7).
a carbon fiber/expoxy layer to withstand higher pressure and an In conclusion, the main steps of the current adopted approach
external glass fiber/epoxy layer, as an impact revelatory in addition can be gathered as following:
to reinforcing the structure strength.
More precisely, these specimens were cut longitudinally from i. perform experiment to build database of the collected
the cylindrical part of the vessel (i.e. parallel to the cylinder axis), signals,
as it can be seen in Fig. 2. ii. save this database in a server or the disk of the microcom-
As a first step of validating the proposed technique, it is suitable puter to be used for post-processing
to work only with composite layers and don’t take into account the iii. run the developed tool. The next section is devoted to detail
influence of the liner. The liner, as a polymer, is avoided for the this task.
moment because it has a material behaviour and damages mecha-
nisms completely different from that of composites. For a raison of 4. Algorithmic implementation
practicality, the liner was removed from the vessel before starting
the cut. Then, the abrasive Water Jet technology (AWJ) was used to A signal processing algorithmic tool has been developed under
extract these specimens. Due to the heterogeneous nature of com- the MatlabÒ commercial platform. The main purpose of this tool
posite materials and the large thickness of the vessel, conventional is extracting relevant features from the acquired signals database.
machining techniques are not suitable and may generate further In the current process, the collected data are stored in SQLite for-
damage. AWJ technology was then chosen (Fig. 3) thanks to its mat files. The volume of this file is around 3 Go. This database is
specific advantages (e.g. lack of thermal damage, low tool wear, resource consuming and their process will generate intermediate
small cutting forces and high productivity) [22]. data leading to memory overloading, to avoid this problem a mem-
 S. Yaacoubi et al. / Composite Structures 183 (2018) 338–346 341

Fig. 2. Specimen are cut from the cylindrical part of the vessel.

Fig. 3. Specimens extraction using AWJ technology.

characteristic is very important to understand the energy distribu-

tion as a function of frequency, group velocity, and mode [23].
Wavelet transform appears to be a powerful tool to access time-
frequency domain [24]. Whether continuous or discrete, this tool
was largely used in composite [25–28] and metallic [29] to inves-
tigate fracture mechanisms. The intensity of the energy in particu-
lar frequency ranges was then used to identify the damage. In this
work, the approach is different, since a frequency feature was
defined from Continuous Wavelet Transform (CWT) to sort the
maximum energy which corresponds to the predominant mode.
As explained before, since a relationship can be established
between mode and source orientation, this analysis is a key advan-
tage to trace back the source identification.
The proposed wavelet algorithm is the CWT of a temporal AE
signal. It is a time joint frequency representation that uses a win-
dow in form of a short time wavelet to fit the signal shape. The
use of the wavelet function ensures a good time and frequency res-
olution helping hence to overcome the limits of the Gabor-
Fig. 4. Location of WD sensors to ensure mode separation.
Heisenberg uncertainty [30]. In addition, this wavelet can be scaled
ory mapping operation has been used to create a virtual memory (i.e. stretched or dilated) to cover all the frequencies contained in
RAM (Random Access Memory). the signal. Therefore, the scale is an important parameter in the
calculation of the CWT. Mathematically, the CWT of a temporal
4.1. Wavelet algorithm signal x(t) is defined by:
Z  
1 ts
The non-stationary of AE signals has prompted researchers to CWT s;s ðxÞ ¼ pffiffi xðtÞx dt ð1Þ
s s
analyse the time-frequency domain. It has been proven that this
342 S. Yaacoubi et al. / Composite Structures 183 (2018) 338–346

Fig. 5. Sensors distribution on the tested coupon during the tensile test. Si with (i = 1, 2, 3 and 4) correspond to the sensors used in this experiment.

Table 1
Features of acoustic emission equipment and acquisition parameters. of time and frequency resolution. As the CWT produces initially a
Hardware Vallen AMSY-6 time-scale coefficients, the scales must be converted to frequen-
Software Matlab routine cies. This can be achieved using the following formula:
Amplifiers Vallen AEP3N
Fc
Amplification [dB] 34 Fs ¼ ð2Þ
Discrimination time [ms] 0.2 sD
Rearm time [ms] 0.4
Frequency range [kHz] 20–650
where s is a scale, D is the sampling period, Fc is the center fre-
Sample rate [MHz] 5 quency of a wavelet in Hz, Fs is the pseudo-frequency correspond-
Sensors S1 = R15; S2 and S3 = WD; S4 = R15 ing to the scale s in Hz.
Threshold [dB] 40 An example of a CWT of a temporal AE signal is illustrated in
Fig. 5. In this case, the selected mother wavelet is the complex
morlet with a center frequency of 5 MHz. The result shows that
the energy contained in the signal arrived in a time between 50
and 150 ms. Besides, the maximum energy has a frequency located
between 100 and 200 kHz. These characteristics will be used after-
ward to identify the type of the main damage that has been
occurred within the structure.
All calculations are carried out by means of a Dell computing
station, which has the following characteristics:

RAM 8 GB
CPU Quad CPU 2,67 GHz
Hard disk 232 GB
GPU 256 MB
System Windows 7, 64 Bits

4.2. Memory mapping to save CPU time

Fig. 6. CWT of a MAE temporal signal using the complex morlet as a mother The main purpose of the memory mapping operation is to cre-
wavelet. ate a virtual RAM memory; this can be done by assigning a file
from the hard drive (a software resource) to a physical memory
(hardware resource). The file is seen by the operating system as
where x is the mother wavelet, s is the scale, s is the translation an extension of RAM memory. Thus, the RAM memory length will
time and ⁄ is the complex conjugation. be extended.
The choice of an optimal mother wavelet depends on the sig- The processing starts by connecting the signals SQLite database
nal’s shape. Different families of mother wavelet exist namely: to Matlab with the mksqlite API. Then, a correction data operation
Daubechies, Morlet, Haar, Gaussian, etc. Consequently, different is applied in order to eliminate pre-trigger time. After that, the
tests must be conducted to select the most suitable one in terms magnitudes are quantified on 16 bits. This allows us to select the
 S. Yaacoubi et al. / Composite Structures 183 (2018) 338–346 343

Fig. 7. Flowchart of the global algorithm used in the MAE, including the memory mapping technique.

elementary memory length of the virtual memory and reduce the gation time and main frequency) for each signal (see Fig. 8a). Sub-
calculation time linked to the floating memory. sequently, an original three dimensions plot (T, T, F) has been
The ‘‘mat” file of the virtual memory is created using the FLAG- carried out in order to identify damage mechanisms (see Fig. 8b).
v7.3 that allows us to perform a data compression as soon as the In the best knowledge of the authors, this king of presentation is
file’s size exceeds 2 Go. The function ‘‘matfile” is used to link the novel.
virtual memory’s file to a MATLAB variable in writing mode so that From the performed tensile test (specimen scale), a database of
it can directly access the mat-file stored on the hard drive without about 3 Go was generated. The implemented algorithm is applied
changing the whole memory variables. on it. Two cases (without and with memory mapping) are studied
This last step is necessary to load the data on the virtual mem- in order to assess more accurately the benefit of this technique.
ory. It was found that for the current configuration, the package The table below represents a comparison between the process-
loading of 1000 signals can be done without overloading the mem- ing time (CPU) in the two cases: without and with memory map-
ory. Consequently, we can process by blocs of Big Data sets whose ping, for different number of signals. As it can be remarked the
size doesn’t exceed the memory capacity. CPU is lower when using the memory mapping, and this is inde-
As soon as the loading is done we will use the Matlab function pendent on the number of signals. Additionally, the higher the
‘‘memmapfile” to map the file (or a part of it) to a Matlab variable number of signals, the more efficient is the algorithm: for 2000 sig-
located in the memory. This facilitates the drive access, for nals, the CPU time wined is 12% however, it is 54% (i. e. more than
resource consuming data which can’t be stored on the RAM mem- half) for the case of 5000 signals.
ory or take too much time to load. Then, the CWT is directly Moreover, when the number of signals is increased (for example
applied to the signals on the virtual memory in order to calculate 10 000), the calculation cannot be achieved without memory map-
the propagation and the arrival times. ping. Using memory mapping, computing is finished and the CPU
time is 1351 s, which is less than 1422 (i.e. two times 711, the
5. Results CPU time required in the case of 5000 signals).
For the ease of the reader, Table 2 is transformed in a barogram
The said tool is based inter alia on CWT, and allows calculation representation, as it can be seen in Fig. 9 (left). The same work has
and editing the coordinates of the mode energy maximum (propa- been done for the memory. Fig. 9 (right) presents the correspond-
344 S. Yaacoubi et al. / Composite Structures 183 (2018) 338–346

Fig. 8. a) Frequency and propagation time location of the maximum WT coefficient for each waveform. b) A three dimensions TTF plot during tensile test. Red continued curve
corresponds to the applied loading. Max is the maximum of the mode energy of the signal at hand. (For interpretation of the references to colour in this figure legend, the
reader is referred to the web version of this article.)

Table 2 memory mapping is not applied. In conclusion, this algorithm

Processing time needed to perform calculation in both cases: with and without has improved significantly data processing.
memory mapping. The symbol 1 indicates that, due to insufficiency in memory,
calculation was not finished.
Finally, Fig. 8b obtained with 5000 signals can be compared
with Fig. 10 (see below) obtained with 35,000 signals. It should
Signals number Algorithm be noted that without memory mapping, it was not possible to pro-
Processing time (s) cess more than 5000 signals. So, it was not possible to obtain a den-
Without memory mapping With memory mapping ser figure with TTF population. Thanks to memory mapping
2000 307 269
algorithm, a much denser figure can be obtained such the result
5000 1526 711 shown in Fig. 10.
10,000 1 1351 To conclude, it was found that the time and memory consuming
35,000 1 5516 of waveforms analysis slow features extraction. Logically, extrapo-
lation to a real structure will lead certainly to a very big file. The
reduction of the consumed memory and computing time is thus
required, to allow establishing an automated procedure for real
ing result. As it can be seen the memory needed for 10,000 signals time inspection of composites structures via MAE. A novel algo-
is roughly identic to that needed for the processing of 5000 signals, rithm using memory-mapping solution was then proposed. In
when the memory mapping is used. It can also be remarked that few words, it consists on file mapping on disk to a range of memory
the memory required for the analyses of 35,000 signals is not much addresses within the Matlab address space, accelerating then file
bigger than the one needed for the case of 5000 signals, when reading and writing.

Fig. 9. Comparison of performance in processing time (left) and memory (right) of the developed algorithm for damage mechanism identification: red (without memory
mapping) and green (with memory mapping). (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)
 S. Yaacoubi et al. / Composite Structures 183 (2018) 338–346 345

Fig. 10. A three dimensions TTF plot during tensile test (35,000 are used to plot this figure).

6. Conclusions References

This paper has presented a case of study using MAE technique, [1] http://www.horizonhydrogeneenergie.com/.
[2] http://www.hypactor.eu/.
where the acquired data by only four sensors is enormous. The ini- [3] Yaacoubi S, Chehami L, Aouini M, Declercq NF. Ultrasonic guided waves for
tial aim was to develop signal processing method in order to reinforced plastics safety. Reinf Plast 2016. http://dx.doi.org/10.1016/j.
improve the MAE reliability, when applied on composite struc- repl.2016.12.002.
[4] The markets: Oil and gas; 2011. <http://www.compositesworld.com/articles/
tures. The encountered problem is the saturation of the computer the-markets-oil-and-gas-2010>.
memory. An algorithmic tool has been developed and applied with [5] Advances in FRP Composites in Civil Engineering: Proceedings of the 5th
success on this case of study. This tool allowed performing a novel International Conference on FRP Composites in Civil Engineering, (CICE 2010),
Sep 27–29, 2010, Beijing China. Ye L, Feng P, Yue Q, editors. ISBN 978-7-302-
data processing which aims to present information in 3 dimen- 23910-9 (Tsinghua University Press, Beijing), ISBN 978-3-642-17486-5
sions, which are: time, time, and frequency. This novel presenta- (Springer, Heidelberg Dordrecht London New York), e-ISBN 978-3-642-
tion helped inter alia determining, under some circumstances, 17486-5.
[6] Gorman MR. Modal AE analysis of fracture and failure in composite materials,
damages mechanisms. In turn, the based mapping-data algorith-
and the quality and life of high pressure composite pressure vessels. J Acoust
mic tool permitted: Emission 2011;29.
[7] Jollivet T, Peyrac C, Lefebvre F. Damage of composite materials, 5th Fatigue
Design Conference. Fatigue Design 2013:1–13.
- processing Big Data files, which were impossible without this
[8] Surgeon M, Wevers M. Modal analysis of acoustic emission signals from CFRP
tool laminates. NDT & E Int 1999;32:311–22.
- win computing time [9] Garret KW, Bailey JE. Multiple transverse fracture in 90 cross-ply laminates of a
- win memory. glass fibre reinforced polyester. J Mater Sci 1977;12:157–68.
[10] Haselbach W, Lauke B. Acoustic emission of debonding between fibre and
matrix to evaluate local adhesion. Compos Sci Technol 2003;63:2155–62.
The achieved result and the adopted approach will be followed [11] Gorman MR, Ziola SM. Plate waves produced by transverse matrix cracking.
to be generalized for enhancing damage discrimination in compos- Ultrasonics 1991;29(3):245–51.
[12] Ramirez-Jimenez CR, Papadakis N, Reynolds N, Gan TH, Purnell P, Pharaoh M.
ite, large, full-scale and complex structures. This is a first step Identification of failure modes in glass/polypropylene composites by means of
towards Big Data solutions to manage as well as possible MAE data the primary frequency content of the acoustic emission event. Compos Sci
in Structural Health Monitoring of large structures. Technol 2004;64:1819–27.
[13] Li X, Ramirez C, Hines EL, Leeson MS, Purnell P, Pharaoh M. Pattern recognition
of fiber-reinforced plastic failure mechanism using computational intelligence
Acknowledgements techniques. In: Neural Networks, IEEE World Congress on Computational
Intelligence; 2008. pp. 2340–2345.
[14] Bohse J, Chen J. Acoustic emission examination of mode I, mode II and mixed-
The accomplishment of the work presented in this paper started
mode I/II interlaminar fracture of unidirectional fiber-reinforced polymers. J
under the project H2E which aimed to contribute to the emergence Acoust Emission 2001;19:1–10.
of a sustainable and competitive hydrogen energy sector in France [15] Dahmene F, Yaacoubi S, El Mountassir M, Bendaoud N, Langlois C, Bardoux O.
On the modal acoustic emission testing of composite structure. Compos Struct
and in Europe, and finished in the frame on the European project
2016;140:446–52.
Hypactor. In both projects, AE testing and monitoring techniques [16] Abdessalem B, Drai R, Dris L, Farid Ch. Enhancement of Phased Array
are very involved. Authors thank H2E partners and more particu- Ultrasonic Signal in Composite Materials using TMST Algorithm. In:
larly Stelia Composites and Air Liquid for their technical support Proceedings of the 2015 ICU International Congress on Ultrasonics, Metz,
France, Physics Procedia, 2015. pp. 488–491.
and availability. The specimen used in this paper study is extracted [17] Li Chuan, Pain Damien, Wilcox Paul D, Drinkwater Bruce W. Imaging
from a vessel provided by Stelia Composites. The authors are grate- composite material using ultrasonic arrays. NDT & E Int January
ful for BPI France for co-funding H2E project. Hypactor has 2013;53:8–17.
[18] Assunção Marcos D, Calheiros Rodrigo N, Bianchi Silvia, Netto Marco AS, Buyya
received funding from the European Union’s Seventh Framework Rajkumar. Big Data computing and clouds: trends and future directions. J
Programme (FP7/2007-2013) for the Fuel Cells and Hydrogen Joint Parallel Distrib Comput 2015;79–80:3–15.
Technology Initiative under grant agreement n° 621194. [19] Dahmene F, Yaacoubi S, El Mountassir M. Acoustic emission of composites
structures: story, success, and challenges. Phys Proc 2015;70:599–603.
346 S. Yaacoubi et al. / Composite Structures 183 (2018) 338–346

[20] Joseph L. Rose. Ultrasonic guided waves in solid media, ISBN 978-1-107- [26] Komai K, Minoshima K, Shibutani T. Investigations of the fracture mechanism
04895-9, Cambridge University Press 2014, 512p. of carbon/epoxy composites by AE signal analyses. JSME Int 1991;34.
[21] José Martínez-Jequier, Antolino Gallego, Elisabet Suárez, Francisco Javier [27] Suzuki M, Kida S, Shimbo M, Miyano Y. Effects of fiber content on fracture
Juanes, Ángel Valea. Real-time damage mechanisms assessment in CFRP mechanisms of short fiber reinforced PET composites. In: Proc. 8th Int. Conf on
samples via acoustic emission Lamb wave modal analysis. Compos B 2015;68 Composite Materials, vol. 3; 1991.
(3):317–26. [28] Suzuki M, Nakanishi H, Iwamoto M, Jinen E. Application of static fracture
[22] Alberdi A, Suárez A, Artaza T, Escobar-Palafox GA, Ridgway K. Composite mechanisms to fatigue fracture behavior of class A-SMC composite. In: Proc.
cutting with abrasive water. Proc Eng 2013;63:421–9. 4th Japan-US Conf. on Composite Materials; 1988. pp. 297–306.
[23] Hamstad MA, Gallagher AO, Gary J. A wavelet transform applied to acoustic [29] Piotrkowski R, Gallego A, Castro E, Garcia-hernandez MT, Ruzzante JE. Ti and
emission signals: part 1: source identification. J Acoust Emission 2002;20. Cr nitride coating/steel adherence assessed by acoustic emission wavelet
[24] Qi G. Wavelet-based AE characterization of composite materials. NDT Int analysis. Nondestruct Test Eval Int 2005;38:260–7.
2000;33:133–44. [30] Ülker-Kaustell M, Karoumi R. Application of the continuous wavelet transform
[25] Loutas TH, Kostopoulos V, Ramirez-Jimenez C, Pharaoh M. Damage evolution on the free vibrations of a steel-concrete composite railway bridge. Eng Struct
in center-holed glass polyester composites under quasi-static loading using 2011;33(3):911–9.
time-frequency analysis of acoustic emission monitored waveforms. Compos
Sci Technol 2006;66:1366–75.