COMPO

w
x ME FEATURE
From Materials Evaluation, Vol. 68, No. 8, pp: 880–886.
Copyright © 2010 The American Society for Nondestructive Testing, Inc.

SITES
Compression After Impact
Load Prediction in Graphite
Epoxy Laminates
by Anthony M. Gunasekera

A
s technology evolves, extensive resources are being utilized in the
research and development of new composites to replace metals in
various structures in many industries. Composites have the advan-
tages of better corrosion resistance, better fatigue properties and
higher strength, as well as a higher stiffness-to-weight ratio than most metals.
However, with increasing advantages there are also disadvantages. The
sudden failure of composite structures is one disadvantage that has been crit-
ically analyzed and has become a vastly intricate area of research, as there
are many varying factors that play major roles.
If laminates are to be used as thin outer surfaces or as the skin of an
aircraft, these composites will be subjected to environmental conditions and
other occurrences during flight that can alter their properties. Low energy
impact damage (8 – 20 J), such as small bird strikes, hail, dropping of tools
on a wing section, or runway stones and debris, can cause different modes of
failure, including internal or surface delaminations and matrix cracking within
the composite. Although this damage may be barely visible from the exterior
of the structure, it can cause an extensive reduction in the compressive
performance or residual strength. This type of impact damage is referred to as
barely visible impact damage (BVID), which further weakens materials that are
already much weaker in compression than in tension. In general, composites
are brittle and provide little to no warning before failing. BVID causes compos-
ites to fail or buckle under much lower compressive loadings than usual.
Therefore, implementation of a reliable technique to determine the compres-
sive residual strength of a composite under compression would greatly reduce
the risk of sudden failure in a composite structure while in service (Dion,
2006; Hill et al., 2008; Karl, 2006; Nguyen, 2005; Walker and Hill, 1997).
Identification and monitoring of these structural changes on an aircraft during

AUGUST 2010 • MATERIALS EVALUATION 881

ME FEATURE w
x compression in laminates
A few hits of higher amplitude were visible,
but they were relatively sparse and suspected
to be occurrences of microdelaminations
in-flight conditions is crucial, as it allows designers to
know the limitations of their products.
Acoustic emission testing (AE) has been success-
fully proven on a broad scale in distinguishing and
classifying failure mechanisms based on the acoustic
signatures of most fiber-matrix composites (ASNT,
2005; Dion, 2006; Hill et al., 2008; Nguyen, 2005).
Although it is challenging to accurately distinguish
between coupled failure modes, failure mechanisms
can be distinguished by the acoustic signature emitted
during composite failure. Failure mechanisms tend to
emit sounds in a unique and consistent manner. While
the ultimate failure loads between two different cases
may differ entirely, the acoustic characteristics
maintain similar failure mechanism trends. These
trends can be exploited by artificial neural networks
(ANNs), computer algorithms that can easily handle
complex multivariate data sets. ANNs have been
shown to accurately classify failure mechanisms based
on data trends and to make predictions for compos-
ites after impact residual compressive strength (Dion,
2006; Hill et al., 2008; Nguyen, 2005).
Two investigations regarding ANNs were conducted
by fabricating and impacting graphite/epoxy coupons
at various energy levels, simulating low velocity impact
damage. The plates were then placed in a material
compression testing machine to test the compressive
qualities of the specimens. AE transducers were used
to listen and collect discontinuity growth activity data
from the coupons under compression. Additionally,
ambient noise from the testing environment was
recorded to properly filter it out of the data set.
First, a Kohonen self-organizing map (SOM) was
constructed and optimized to separate the AE data
into the various failure mechanisms experienced by
composite laminates undergoing compression. It was
hoped that by quantifying the failure mechanisms,
more accurate ultimate strength predictions could be
made. A Kohonen self-organizing map is an unsuper-
vised, competitive, two-layered artificial neural
network (ANN), which is incredibly useful for data clas-
sification applications. This mapping system is based
on the human brain, which uses special organization
to classify certain objects or organisms into separate
groups.
882 MATERIALS EVALUATION • AUGUST 2010
Secondly, two backpropagation neural networks
(BPNNs), which weigh the amplitude distributions
from AE data to arrive at accurate compressive
strength predictions, were constructed and optimized.
The first BPNN analyzed data had not been filtered for
noise, and were not classified into mechanisms using
a SOM. The second BPNN analyzed data that the SOM
classified into distinct failure mechanisms and were
used to eliminate noise. This way, comparisons could
be made as to the effectiveness of SOMs in data clas-
sification for prediction purposes.
A worst case error of 5 to 15% between the
predicted load and the actual test coupon failure load
was desired for this analysis. If these neural networks
could be successfully trained using AE test data, it
could then be possible to use this same technique
during in-flight conditions. AE transducers could be
mounted on the skin, panels or other structures and
used to listen and collect discontinuity growth data.
These data could then be fed into previously trained
neural networks that were trained on test samples and
accurately predict the failure modes and final residual
compression load of the structure. This procedure
could be done continuously in-flight and on the
ground, saving companies money as it would reduce
maintenance inspections and unnecessary changes in
parts before the actual failure life was reached.
Material and Testing Techniques
A graphite bidirectional woven prepreg cloth was used
to fabricate thirty-four 101.6  152.4 mm (4  6 in.)
laminate coupons. The prepreg had a manufacturing
discontinuity due to misaligned fibers that were
present in the weave. It was thought that this could
prove to be a problem, as fibers misaligned in the axis
of compression could buckle prematurely, increasing
the variability in the compressive strength of the
coupon. However, since all of the coupons were manu-
factured in the same way with the same discontinuity,
it became less of a concern. The thickness of prepreg
tape suggested that a 24 ply lay-up be used to
construct the coupon laminates.
Impact Testing and Imaging
BVID was created using a calibrated impactor. The
impactor was configured with a blunt 12.7 mm (0.5
in.) hemispherical tup. The tup simulated a low 101.6 mm
velocity impact such as a tool dropping on the (4 in.)
coupon. The impactor was set to impact at energies Compression fixture 50.8 mm
(2 in.)
ranging between 8 and 20 J. An ultrasonic C-scan and 150 kHz Resonant AE
X-ray image were also performed to analyze the effects transducer

of the impact damage on the coupon.

38.1 mm
Compression Testing and Acoustic Emission (1.5 in.)

A material compression testing machine was used to

Graphite/epoxy test coupons
compress the specimens. Compression after impact
152.4 mm
testing (CAI) is an ASTM standard test that is used to (6 in.)
test the compressive strength of aerospace materials Impact damage

after they have been impacted with low or high energy

impacts. The CAI test fixture was used to secure the 38.1 mm
coupons for compression (Figure 1). After securing the (1.5 in.)

coupon in the compression after impact fixture, the

specimen was placed in the compression machine
and the coupons were compressed at a constant rate
of 296.548 N/s (66.667 lbf/s). Upon failure, the
applied force dropped from the maximum load to zero
and the machine was stopped. Acoustic emission data
were recorded during the compression test using an
acoustic emission analyzer. This handheld device
employs two 150 kHz transducers to collect acoustic
emission data. The transducers were hot-melt glued
(used in the industry to mount transducers) to the
specimens on the centerline from the top and bottom
edges of the coupon. The transducers were also taped
to the compression fixture to avoid damage in case
buckling failure caused them to violently separate from
the coupon. The acoustic emission analyzer also
allowed for a parametric input of the load from the
compression machine so the final compressive load
could be obtained. This was also used to visualize the
compressive life of the coupon and analyze how load
changes with failure mechanisms.
Figure 1. Compression test setup.
being hot-melt glued to the laminates, but with no
compressive loads applied and the compression
machine crosshead stationary. The compression
machine was turned on and recordings were
conducted for approximately the same duration as a
typical compression test, ranging from 180 to 300 s.
Once this noise signature was separately analyzed, it
was clearly seen in acoustic emission data taken from
a specimen test. Figure 2 shows this low amplitude
noise with specimen data in a hits versus amplitude
graph.

Data Filtering and Noise Analysis

A software code was created to eliminate erroneous 300
data points and filter the data efficiently. Hits that Low amplitude noise (30 – 36 dB)
250
recorded values of zero (extremely weak hits) for
energy, duration or rise time were all removed first. 200
Long duration hits greater than 9800 µs were also Specimen amplitude distribution
Hits

150
removed (these being thought to be rubbing noise band (34 – 100 dB)
from the sides of the compression fixture). 100
Additionally, due to the wide variety of random noise 50
signals that occur during the progression of any type
of testing, a proper noise analysis was important. In 0
30 40 50 60 70 80 90 100
order to truly understand the acoustic signature of
ambient noise and thus remove it, five recordings of Amplitude (dB)
the quiescent testing environment were taken. All of
these recordings were captured using the pocket Figure 2. Hits versus amplitude plot showing the difference between noise and
acoustic emission transducer, with the transducers specimen data.

AUGUST 2010 • MATERIALS EVALUATION 883

ME FEATURE w
x compression in laminates

Figure 3. Laminate with 15 J BVID, displayed using C-scan and X-ray imaging.

Results and Discussion
Effects of Impact Testing
Figure 3 shows a laminate coupon that was impacted
at 15 J. To the naked eye, it is difficult to see any
damage. However, when an ultrasonic C-scan and X-
ray image are taken of the same coupon, the damage
is visible. Longitudinal and transverse cracking is
clearly seen at the site of impact along the weave of
the composite.
Effects of the Compression and Noise Analysis
Once testing was completed and all of the failure
loads were acquired, Figure 4 was plotted to visualize
30 000
28 000
Standard data
Compressive failure load (lbf)
how the coupons were affected by the varying impact
energies. It can clearly be seen that as impact energy
increases, the residual CAI failure load of the
laminates is affected in a parabolic fashion. A signifi-
cant amount of variability is observable, but this is to
be expected, given the nature of composite materials
and the manner of fabrication. The parabolic nature of
the compressive failure load was expected, because
as impact energy increases, the type of impact
damage changes from barely visible to a high velocity
impact that penetrates the surface. This penetration
creates a circular indentation that is thought to inhibit
slip between plies. This effect would then increase,
rather than decrease, the compressive strength. The
latter line is the tolerance interval, the B-basis
allowable for the composite coupons.
The first graph obtained from acoustic emission
data to be discussed will be amplitude versus time, as

26 000 Best fit parabola shown in Figure 5 for Coupon 23A. The time scale is
24 000
B-basis allowables from the start of the test all the way to failure. Note
22 000
that in the first half of the test, the vast majority of the
hits occur in a range thought to be consistent with
20 000
matrix cracking. A few hits of higher amplitude were
18 000
visible, but they were relatively sparse and suspected
16 000 to be occurrences of microdelaminations. Matrix
14 000
cracking, as this range is assumed to be, occurs evenly
across the time scale. At the beginning of the second
12 000
half of the test, hits in the range consistent with
10 000
8 9 10 11 12 13 14 15 16 17 18 19 20
delaminations become more frequent. Additionally,
Impact energy (J) hits in the range consistent with fiber breaks also
begin to occur, becoming more prevalent as the test
proceeded to failure.
Figure 4. Failure load versus impact energy plot showing parabolic relationship.

884 MATERIALS EVALUATION • AUGUST 2010

Failure Mechanism Classification through Kohonen
Self-Organizing Maps
100
Macro delaminations
After considerable trial and error, it was determined micro/macro fiber breaks (70 – 100 dB)
that the SOM would best classify the various failure 90
Micro delaminations (55 – 72 dB)
mechanism signals using the amplitude and average 80

Amplitude (dB)
frequency parameters. Alternative analyses were
70
conducted using amplitude, energy and duration, but
the best results were produced using the former 60
combination. Using amplitude, average frequency,
50
energy and other separate parameters during testing
resulted in the same type of plots acquired from the 40
Matrix cracking (34 – 62 dB)
network tested on very few parameters. Values from
30
each of the 34 samples were then used to make indi- Start of test failure
vidual testing files, which revealed slightly varying Time (µs)
outputs for each coupon. As for the number of
Kohonen layer processing elements, training iterations
were performed with as few as two and as many as Figure 5. Amplitude versus time graph showing when certain mechanisms are
experienced.
twenty processing elements. The number of processing
elements approximate how many groups or clusters
the data will be divided. Graphs were created for every
300
variation of processing elements in the network. Noise
The number of Kohonen layer processing elements Matrix cracking Mech1
250
will equal (2n + 1)  (the number of choices [or Mech 2
Mech 3–4
clusters of data the network is given for classification]). 200
This SOM was configured to assign X and Y coordi-
Hits

nates to input vectors with similar amplitudes and 150

average frequencies. The number of Kohonen
processing elements determined the number of 100
distinct amplitude and average frequency ranges to Micro delaminations
which an input vector could be classified. In a complex 50 Macro delaminations
Low amplitude micro/macro fiber breaks
classification problem such as composite failure mech- noise
0
anisms, a SOM by itself cannot determine the ideal 30 40 50 60 70 80 90 100
failure mechanism distributions; rather, it provides
Amplitude (dB)
ranges of similar data, which must be interpreted to
determine plausible mechanism distributions.
For this research, the best results were obtained Figure 6. Amplitude histogram of data classified by a SOM.
using 14 Kohonen layer processing elements (or data
ranges), then combining the output into four distinct
groups: low amplitude noise, matrix cracking, microde-
100
laminations, and a combination of macrodelamina-
Macro delaminations
tions and fiber breaks. These results are shown in a 90
micro/macro fiber breaks
hits versus amplitude and an amplitude versus
80
average frequency graph for the coupon tested in
Amplitude (dB)

Figures 6 and 7. All this was after an iterative process 70

Micro delaminations Noise
in which the 14 separate ranges were combined into
60 Mech 1
four plausible failure mechanisms. Specifically, the Mech 2
humps thought to be external noise, microdelamina- 50 Mech 3–4
tions, and the combination of macrodelaminations
and fiber breaks were left unchanged – they are 40
Matrix cracking
possibly unique ranges that were output by the SOM. 30
The other 11 ranges of data were combined to 0 50 100 150 200
Low amplitude noise
produce the dominant hump denoted as matrix Average frequency (kHz)
cracking. This suggests that within the mechanism of
matrix cracking, there may be 11 different modes of Figure 7. Amplitude versus average frequency graph of data classified by a SOM.

AUGUST 2010 • MATERIALS EVALUATION 885

ME FEATURE w
x compression in laminates
cracking or possibly 11 different specimen reso-
nances, each emitting slightly differing acoustic signa-
tures. It could also be possible that, as the processing
elements were increased, the SOM itself found unique
characteristics in the data of low amplitude noise,
microdelaminations, and the combination of macrode-
laminations and fiber breaks, and could only further
divide the data considered to be matrix cracking into
clusters.
After-Impact Load Prediction Using Backpropagation
Neural Networks
In order to predict the CAI failure load in the coupons,
BPNNs were constructed and optimized using the
same software as for the SOM. In the training phase,
the networks were trained by normalizing 24 of the
coupons’ amplitude distributions (containing data up
to 50% of the failure load) as the inputs. The known
failure load for each of these coupons was also
entered into the network. The network then updated
the weights between the processing elements until the
output failure load came within a certain percent error
of the actual failure load, at which point the ANN was
considered to be trained.
In the testing phase, the networks were tested
using the amplitude histograms from the other ten
samples, with the corresponding failure loads being
predicted. Two general BPNN architectures were used,
and various network parameters were optimized for
each configuration until the most accurate predictions
were reached. In the first configuration, the input layer
contained 72 processing elements, one for each
amplitude decibel level between 30 and 100 dB, plus
one for the actual failure load (during the training
phase only). The number of hidden layer processing
elements was varied from 7 to 15 in order to find the
optimum value, and the output layer consisted of a
single processing element, the predicted CAI failure
load. At this point the input data were not classified
and therefore still contained external (specimen/test
fixture rubbing) noise.
For the second configuration, the input layer
contained 214 processing elements. Amplitude distri-
butions from each individual mechanism were used as
inputs. The data classified as external noise were
removed, leaving three amplitude distributions: one
for matrix cracking, one for microdelaminations, and
one for the combination of macrodelaminations and
fiber breaks. As before, the actual failure load was
also an input (for training), the number of hidden layer
processing elements varied between 7 and 15, and
886 MATERIALS EVALUATION • AUGUST 2010
the sole output processing element was the predicted
failure load. After the optimization process was
completed for each configuration, the lowest
maximum prediction error was found to be 15.89% for
the unclassified data. On the other hand, for the clas-
sified data with the noise removed, the lowest
maximum error was reduced to –11.65%. After
approximately 65 more iterations attempting to
optimize the minor parameters for the second configu-
ration, the maximum error was ultimately reduced to
–11.53%.
Conclusion
In conclusion, using AE and ANNs to classify failure
mechanisms and predict compressive failure loads for
this research proved highly successful. The combina-
tion of these techniques could be used on aircraft to
predict the failure loads associated with impact
damaged parts without knowing a priori the level of
impact damage. Moreover, these predictions could be
made in-service using AE. w x
AUTHOR
Aerospace Engineering Department, Embry-Riddle Aeronau-
tical University, 600 South Clyde Morris Road, Daytona
Beach, FL 32114.
ACKNOWLEDGEMENTS
The author would like to thank Michael W. Langdon, William
D. Rice and Eric v.K. Hill for their assistance in the research,
writing and editing of this paper.
REFERENCES
ASNT, Nondestructive Testing Handbook: Acoustic Emission
Testing, third edition, Vol. 6, Columbus, Ohio, American
Society for Nondestructive Testing, 2005.
Dion, S.A.T. “Neural Network Burst Pressure Prediction in
Composite Overwrapped Pressure Vessels from Acoustic
Emission Data,” MSAE Thesis, Daytona Beach, FL, Embry-
Riddle Aeronautical University, 2006.
Hill, E.v.K., Y. Zhao, T. Ebert, J.V. Kay and G.K. Lewis,
“Compression After Impact Strength Prediction in Compos-
ites Using Acoustic Emission and Artificial Neural
Networks,” American Institute of Aeronautics and Astronau-
tics, 49th AIAA/ASME/ASCE/AHS/ASC Structures, Structural
Dynamics and Materials Conference, Schaumburg, IL, 2008.
Karl, J.O., “Filtering of Acoustic Emission Data through Prin-
cipal Frequency Component Extraction,” MSAE Thesis,
Daytona Beach, FL, Embry-Riddle Aeronautical University,
2006.
Nguyen, T.D., “Damage Assessment and Strength Predic-
tions in S2-Glass/Epoxy Laminates Subjected to Low Energy
Impact,” MSAE Thesis, Daytona Beach, FL, Embry-Riddle
Aeronautical University, 2005.
Walker, J.L. and E.v.K. Hill, “An Introduction to Neural
Networks: A Tutorial,” First International Conference on
Nonlinear Problems in Aviation and Aerospace, Daytona
Beach, FL, 1997, pp. 667-672.