Академический Документы
Профессиональный Документы
Культура Документы
1 Introduction
Composite materials such as carbon fibre, are continually finding new and ever more
interesting applications in the world of engineering, they are now widely used in the
aerospace industry and are increasing being used in the automotive sector, particularly
in motorsport where every effort to save weight is made.
These new composite materials can be designed to be stronger and lighter than
traditional materials, giving composite materials a significant performance advantage.
Whilst superior in many areas composites are however very susceptible to damage
which is an obvious major drawback. With traditional materials like steel, impact
damage will typically form a dent in the surface of the material and not significantly
affect the performance of the system. In a composite material however the effect is
more severe; because composite materials are made from a combination of fibres and
resin, when it is impacted the fibres become damaged causing a significant weakening
of the material and hence the engineering system built from it [1].
The method by which composite materials are manufactured can also cause
performance issues. Composite parts are typically made by layering sheets of
composite material, called plies, on top of one another. If this process is not properly
executed then multiple issues may arise. Plies could be laid and an overlap formed,
there could be gaps between plies or there could be ply drop off. One of the most
critical flaws that can be formed is known as a delamination this is where the bonding
L. Iliadis, H. Papadopoulos, and C. Jayne (Eds.): EANN 2013, Part I, CCIS 383, pp. 203–212, 2013.
© Springer-Verlag Berlin Heidelberg 2013
204 R. Amali and B.J. Hughes
between two plies has broken down and the two plies have come slightly apart
causing an air gap to form. The occurrence of delaminations is a problem because it
reduces the strength of the material, and is a possible source of further damage
through the life of the part. Delaminations often occur as a result of impact damage.
With such a dramatic uptake of composite usage, particularly in the aerospace
industry, it is important that any defects in servicing material are quickly found. The
detection of these flaws in often carried out through Non-Destructive Testing (NDT).
The use of NDT aims to identify the presence of defects within a material, determining
its location, and in many cases the size and type of defect found. Many different NDT
methods have been developed for composite materials, including radiography,
ultrasound and thermography. The issues here is that despite the developments made in
the use and development of non-destructive testing for composite materials and
structures, few simple, cheap and effective methods exist for the testing of composite
materials. A number of methods have been developed but these are often too complex to
be quickly applied in the engineering workplace [2] [3].
One method that does exist however is the tap test method, which is a commonly
used in-service inspection technique used by many as the first testing method for
composite materials. The tap test can be used to find the presence of many of the
defects present in composite material, namely delaminations. Similarly to finding a
wall stud, the material is tapped using a tool, commonly a coin or small hammer, and
the acoustic responses observed. By noting the changes in acoustic tone it is possible
to use tap testing to find the location of delamination in composite materials [4].
Whilst tap testing is somewhat successful, it typically suffers from a number of
problems, most noticeably the subjective interpretation of acoustic results.
This paper details the research undertaken in the design of a new NDT method that
aims to remove the problem of subjective human interpretation of acoustic tones,
instead using Digital Signal Processing (DSP) combined with Artificial Neural
Networks (ANNs) to perform the interpretation of the obtained test data.
Fig. 1. A comparison of the frequency spectra obtained from the coin tap testing data. The top
graph shows the data from the sample in the undamaged state, with the bottom graph showing
the damaged state results. There is a clear difference between the results obtained.
206 R. Amali and B.J. Hughes
The final phase of the investigation involved the development of classifier that
would be able to correctly determine whether the acoustic data came from the sample
in either the damaged or undamaged state. To do this a 3 layer feed-forward neural
network was created and programmed using the frequency data obtained. The input to
the network would be the 30 average frequency band values for each of the data sets,
with the training target being the corresponding state of the sample, either damaged or
undamaged. For the hidden layer it was decided that 10 neurons would initially be
used. For each of the 8 sound input methods, 5 classifiers were trained to allow for a
more complete evaluation of their performance.
With the classifiers trained it would then be possible to validate their performance.
Throughout the experimental process additional recording had been made in
preparation for this. The new validation data was processed in the same way as before
to get the average frequency band values. The validation data contained 20 sets of
results from the undamaged sample, and 20 from after it was damaged, 40 in total for
each of the 8 input types.
The unseen data sets were fed into the networks and the networks outputs were
compared against the actual classification of the data. The success of the ANN was
measured by the percentage of correct classifications that were made.
makes it difficult for this testing method to be used effectively for real engineering
systems. A different approach however using different microphones and making use
of unsupervised learning could make it a more relevant testing method.
3.1 Methodology
The methodology for this investigation is broadly similar to what was conducted
previously. It is broken down into a 3 phase process, manufacture and data
acquisition, data processing and network design and validation.
In total 3 composite plates were manufactured for this investigation, made from
aircraft grade unidirectional carbon-fibre pre-preg material. Each plate was made
using four 250mm square plies/layers and cured using a heated press. The first two
plates were made to be undamaged so as to give good data for which to train and
208 R. Amali and B.J. Hughes
validate the networks. The other plate was manufactured so that it contained an
artificial delamination; made using a 7mm diameter disc placed in the center of the
plate. This would generate damaged state data that could be used to further validate
performance of the networks.
Fig. 2. The manufacture of the new plates required one of them to contain an artificial
delamination. Above shows how the delamination was formed, by placing a piece of backing
material between the top two plies stopping them from bonding during the curing process.
With the composite plates manufactured the next step dealt with the acquisition of
contact microphones with which to conduct the tests. Two types of contact
microphone were sourced, the first type were the kind that are used to tune musical
instruments, with the second type of being manufactured using piezoelectric discs.
Fig. 3. The second kind of contact microphone was made by soldering together a piezoelectric
disc and 3.5mm audio jack cable. This set up allows the vibration picked up by the disc to be
recorded by plugging the jack into a relevant device.
With the materials acquired it would then be possible to acquire the test data. The
first undamaged panel was placed into the test rig and secured in position, ensuring a
uniform boundary condition was applied to each edge. Each panel had been split into
9 areas in a 3x3 grid. The central area would be used as the sound input location; with
the microphone attached to the plate in one of the other areas. The sample would then
be tested and the acoustic response recorded, after each test the microphone was
moved to a different area, ensuring that a range of readings was recorded.
The sound inputs used to excite the sample would be some of the inputs used in the
previous investigation. White noise imparted into the material using a vibration
speaker had proved to be successful and so that was again used, as was the use of a
chirping sound. Sound impacted made by tapping the sample with a coin and also a
pair of scissors was also used.
The first composite sample was tested using each of the 4 input types detailed. In
total 20 recordings were made for each input type, with the microphone being moved
Detection of Damage in Composite Materials 209
after each individual test. With testing on this sample complete the composite was
replaced with the second undamaged sample and the experiment repeated to gather
yet more acoustic data.
The data obtained from the first undamaged sample would be used to train the
novelty detector and is termed the clean data. The acoustic data obtained from the
second panel termed the undamaged data, is used to help validate the performance of
the novelty detector.
With the undamaged samples tested the experiment moved on to test the damaged
composite sample. The damaged plate was placed in the test rig and tested in two
different ways. Firstly the microphones were placed in the center of the plate directly
over the delamination; the sound was then imparted into the material in one of the
adjacent areas. Theoretically because the bonding of the plies beneath the microphone
is compromised by the delamination, different sounds would be picked up by the
microphone, compared to before. Each sound input type was again used, obtaining 20
readings for each; the process was repeated for the second variant of microphone.
The second method of testing the panel swapped the position of the microphone
and sound input location. The sound would now be generated directly above the
delamination, with the microphone placed in an adjacent area. Again different sounds
should be picked up by the speaker because of the lack of bonding underneath the
sound input location. Both microphones were used with the four different sound input
types, 20 recording were again made. This concluded the testing on the two
undamaged panels.
With the testing on this panel finished the process of data acquisition was
complete. For each microphone and sound input type, acoustic data had been obtained
from the undamaged panels, as well as from the damaged panel, tested in two
different ways.
minimized. The data processing here followed the same methods as used in the
previous experiment, as this had proven to be successful.
The data processing was applied to all of the data that had been collected
throughout the experimental process to generate for each data set, 30 discrete
frequency band magnitudes, these values could be used to describe the profile of the
frequency spectra for each data set.
range. The difficulty with detecting damage using this method is defining what level
of novelty indicates the presence of damage in the system. A system was proposed
where the cumulative mean and standard deviation of the novelty values is use to set
up a pair of limits, if a novelty value exceeds these limits then it is signaled as being
novel and that there is damage in the system. These limits were applied to each set of
results and the overall results analyzed.
Fig. 4. The results from the detector are shown as a novelty trace. It can be seen that for the
undamaged data sets (0-100#) the novelty value is low and within the warning limits, signaling
no damage is present. Once the detector is presented with damaged state data (101-200), the
novelty value increase, crossing the warning limits hence signaling the presence of a
delamination.
4 Concluding Remarks
The research conducted in this investigation is at best rudimentary. Many possible
complications to the testing method are not considered but none the less as a proof of
concept the investigation has been successful, it has been shown that it is possible to
use the idea of novelty detection to identify the presence of delaminations in
composite materials. A range of possible sound input methods have been evaluated
and proved to be successful, two variations of using this technique were testing with
both producing good quality novelty detectors.
This early research shows that there is potential for this simple testing method to be
further adapted and developed to turn it into a fully fledge NDT method. Further work
can be done to investigate other possible methods of imparting sound into the test
pieces, or other possible methods of testing for damage. The potential for detecting
different damage types could also be explored.
Very little attention in the investigation was paid to the way in which the frequency
spectra were processed. It is highly likely that there is a better way of extracting
relevant frequency features that could be used to detect novelty. It is also noted that
the network architecture was also very simple; because the novelty detectors worked
well first time no effort was made to change the network architecture to possibly
improve the results.
Ultimately though what has been done here has been proven successful, it is
possible to use the concept of novelty detection, as well as classification methods, to
detect the presence of damage in composite material. Further work could be carried
out to improve the detections methods making them more accurate and more versatile.
References
1. Hsu, D.K.: Nondestructive Testing of Composite Structures. In: 17th World Conference on
Nondestructive Testing
2. Bray, D.E., Stanley, R.K.: Nondestructive Evaluation – A Tool inDesign, Manufacturing,
and Service. CRC Press (1997)
3. Halmshaw, R.: Non-Destructive Testing. Butterworth-Heinemann (1991)
4. Falk, J.P., Steck, J.E., Smith, B.L.: A Nondestructive Testing Technique for Composite
Panels Using Tap Test Acoustic Signals and Artificial Neural Networks. Journal of Smart
Engineering System Design 5(4), 491–506 (2003)
5. Amali, R., Hughes, B.J.: Development of a novel NDT technique for damage
classifications of composites. In: FRP Bridges Conference 2012, pp. 89–100.
NetComposites, London (2012)
6. Smith, S.W.: Digital Signal Processing. Newnes (2003)
7. Worden, K.: Structural Fault Detection Using a Novelty Measure. Journal of Sound and
Vibration 201, 85–101 (1997)
8. Farrar, C.R., Worden, K.: An Introduction to Structural Health Monitoring. Phil. Trans. R.
Soc. 303, 303–315 (2007)
9. Giurgiutiu, V., Soutis, C.: Enhanced Composites Integrity Through Structural Health
Monitoring. Applied Composite Materials 19, 813–829 (2012)
10. Worden, K., Manson, G.: The Application of Machine Learning to Structural Health
Monitoring. Phil. Trans. R. Soc. 365, 515–537 (2007)