INTRODUCTION
Nondestructive testing (NDT and E) is the application of technical methods
to examine materials or components in terms of their macroscopic or microscopic
compositions, integrity and mechanical properties without impairing their future
usefulness (ASTM International, 1999). This technology
has been widely applied in industrial and civil engineering for new and inservice
health monitoring of materials. Major NDT and E techniques are Xradiography,
thermography, eddy current, acoustic emission and ultrasonic testing. The adoption
of these techniques in an application is highly dependent on the material properties
of the test object and the nature of the flaws (Ng et al.,
2011). Typically, NDT and E techniques are based on the analysis of the
transmitted signals. Among these techniques, ultrasonic testing has always been
prominent in NDE and T as it offers a great advantage in which it is applicable
on all kinds of materials, having high penetration depth and flexible (Shull,
2002). Ultrasonic NDE works on the principal of detecting and interpreting
the time delayed of the reflected ultrasound waves by boundaries (D’Orazio
et al., 2008). The boundaries could be the back surface of the material
or a discontinuity (e.g., crack, porosity or delamination). The maximal amplitude
of ultrasound echoes is used as a means to characterize the nature, size and
orientation of defects. For any kind of material, scattering of ultrasound waves
occurs as ultrasound pulse is induced to the material during testing. In homogeneous
material, the scattering is not apparent and can be neglected (Rubbers
and Pritchard, 2003). However, in the case of coarsegrained structure such
as composite materials, the scattering effect may annihilate the return echoes
and obscure the defect visibility. This signal drowning problem arises when
the wavelength of the reflected ultrasound signal by a defect and the scattering
of the constituent particles in the propagating medium are at the same order
of magnitude. These echoes are distributed randomly in time and known as backscattering
noise. In the preliminary work, the technique of providing high SNR has been
utilized to resolve the echoes associated with large grained and highly attenuated
material (Bilgutay et al., 1981; Kraus
and Goebbles, 1980). Basically, there are two ways to achieve high SNR in
ultrasonic testing: (1) Optimum frequency selection of an acoustic wave for
detecting specific discontinuity and (2) Increasing the ultrasound wavelength.
The latter method also limits the detectability of small discontinuity.
By signal processing means, significant work to increase SNR includes Splitspectrum
Processing (SSP) (Bosch and Vergara, 2008), Expectationmaximization
(EM) algorithms (Benammar et al., 2008), Hilbert
transform (Drai et al., 2000), HilbertHuang
transform (Kazys et al., 2008) and Discrete Wavelet
Transform (DWT) (Liang and Que, 2009). SSP works on
the principal of frequency diversity to identify the actual defects from grain
boundary. This method decomposes the received broadband signal into a number
of smaller frequency bands by means of digital filtering. Later, each band is
inversed Fourier transform to obtain the individual frequency shifted signals.
These individual signals are then worked on amplitude minimization (SSPAM)
or polarity thresholding (SSPPT) to improve SNR (Ericsson
and Stepinski, 2002). The shortcoming of SSP is the application of single
window for all frequencies in the Fourier transform. This yields the same resolution
of the signal analysis at all locations in the timefrequency plane. Benammar
et al., (2008) work showed that the performance of EM algorithm is
comparable to SSPPT for thickness measurement and defect localization but EM
algorithm requires more computation work as deconvolution is carried out in
time domain. Hilbert transform is very sensitive to signal drowned in noise
especially experimental signals (Drai et al., 2000).
HilbertHuang transform offers the adaptive timefrequency decomposition for
the received signal but this method is highly dependent on the material and
defects to be examined. In other words, different material or defects may require
the selection of different sets of intrinsic modes. DWT gives a very good result
to enhance the detectability of discontinuity in ultrasonic signal processing
(Zhong and Oyadiji, 2007; Tang and
Abeyratne, 2000). Wavelet transforms offer an infinite set of possible basic
functions to approximate the measured signal. Dilation and shifting of mother
wavelet enables the access of signal information in different scales. Thus,
this technique offers a great flexibility to signal interpretation. Wavelet
denoising (Tokmakci and Erdogan, 2009) has been widely
used as noise filtering and pattern recognition tool for image processing due
to the advantage of wavelet transform. Many thresholding rules have been suggested
by (Lazaro et al., 2002). The objective of this
study is to develop a new waveletbased multiresolution signal decomposition
method. This method decomposes the received signal into multiresolutional means
by a selected mother wavelet. In each state of scaling, minimization is implemented
inside each frequency band to discard the noisy data. After a few repetitions,
it can be observed that the amplitude of backscattering noise is averaged whilst
the real defect’s signal amplitude becomes apparent. The main advantage
of the proposed method is that threshold rules of the ordinary wavelet denoising
are replaced by the minimization approach for wavelet shrinkage as an improvement
to the more flexible detection algorithms.
DISCRETE WAVELET TRANSFORM
The fundamental principal of wavelet theory is to analyze according to scale
(Graps, 1995). As a function of zero average (Mallat,
1998), a mother wavelet can be dilated with a scale parameter s and translated
by l:
Where, s is the wavelet’s width
and l is the wavelet’s
position. In DWT, a finite energy signal f can be decomposed over the wavelet
orthogonal basis as Eq. 1 to become:
in which each partial sum of f:
brings out the detail variations at the scale of 2^{s}. Recursively
applying Eq. 2 decomposes the signal f into a matrix of sequence
coefficients d_{1}, d_{2},..., d_{M}, a_{M}.
Figure 1 illustrates the wavelet decomposition of signal f.

Fig. 1: 
Twoband wavelet transform for a foursample signal f(t) 
Vice versa, adding up these layers of computed coefficients following the sequence
ultimately resembles the original signal f:
where, a_{M} and d_{M} are the approximation and the detail
coefficients, respectively. D_{M} are representations of signal f at
higher scale as m increases.
WAVELETBASED MINIMIZATION
This section discusses a multiresolution signal decomposition method to improve
the detection of actual defect’s
signal embedded in noise. The approach of minimization based on signal polarity
is used to accentuate the actual defect’s
signal amplitude out of backscattering noise. Figure 2 illustrates
the scheme of the proposed waveletbased minimization signal processing. The
method consists of three successive steps, namely (1) Signal decomposition by
multiple scales and translations, (2) Grain noise averaging and (3) Signal reconstruction.
Firstly, the input noisy signal is decomposed up to N levels of approximation
and detailed coefficients using a selected mother wavelet. Minimization is then
carried out to average the noise amplitude in each frequency band. Lastly, the
altered coefficients are synthesised to resemble the original signal.
The algorithm shown in Fig. 3 summarizes the procedure of
the proposed waveletbased minimization for defect detection. DWT is applied
to separates the input signal X = x_{0}, x_{1},..., x_{n}
into two frequency bands. In this operation, the wavelet coefficients are halved
to become approximations and details by a chosen twochannel symmetric filter
banks (Abdelnour and Selesnick, 2005). Thus, Eq.
2 can be rewritten as:
where, c_{k} and d_{j,k} denotes the approximation and detailed
coefficients, respectively. j is the scaling factor and k is the shifting factor.
Given Eq. 5, two sets of coefficient series (highpass and
lowpass) are generated. It is noted that c_{k} depicts the highpass
coefficients and d_{j,k} depicts the lowpass coefficients. To avoid
the signal overflow, down sampling by two is implemented as stated in step 3
of the algorithm. Step 47 suppresses the small value coefficients that represent
the noise content.

Fig. 2: 
Computation scheme of the proposed waveletbased minimization
signal processing for f(n) 

Fig. 3: 
The waveletbased minimization algorithm 
The value of the amplitude of the processed signal is set to zero when the
data set exhibits polarity reversal in which the zeromean valued presented
by the backscattering noise signal. Step 9 recombines the processed approximation
and detail coefficients back into the original signal by means of inverse DWT.
To compensate the aliasing in step 3, up sampling is carried out by padding
zeroes between the sample data. Step 47 repeats for each iteration of decomposition
(step 2). This algorithm assumes Gaussian distribution, N(0,1) for the backscattering
noise.
EXPERIMENTAL SETUP
To demonstrate the practical capability of the proposed method, the waveletbased
minimization is tested on simulated signals and actual signals from ultrasonic
testing on coarsegrained structured material. The simulated signals are generated
by Field II (Jensen, 1996; Jensen
and Svendsen, 1992), an ultrasound field simulation program.
Simulated ultrasound data: The measured ultrasound signal s(t) can be
expressed as the sum of two components (Song and Que, 2006):
where, y(t) is the acoustic signal and n(t) is the embedded noise. Signal s(t)
is the ultrasound data that is distorted by noise. A typical ultrasonic signal
y(t) was generated using the simulated 1 MHz transducer. White Gaussian noise
was then added to the waveform. Figure 4 shows the noisy received
signal with SNR = 0.1 and 1024 sample length. To construct a fullrange ultrasound
signal in a pulseecho testing setup, back wall echo and the simulated defect’s
echo were added to y(t).
Experimental system: A four layered GFRP laminates ([0^{0}]_{4})
with the dimension of 200x200x3 mm were prepared. The test specimen has Mylarinduced
delamination at the depth of 0.89 mm from the front surface of the specimen.
A single transducer with 1 MHz centre frequency and diameter of 31.24 mm acted
as transmitter and receiver of ultrasonic signals. The ultrasonic testing system
worked as follows: The transducer, working on piezoelectric effect (Diamanti
et al., 2005), transformed electrical pulses into mechanical vibrations
to generate ultrasonic waves (pulse). The transducer was placed onto the GFRP
specimen and the pulse generated was transmitted into the test object (GFRP).
The pulse travelled through the test object and responded to the object’s
a boundary (i.e., voids, cracks or delamination). The returned signal (echo)
was then transformed back into electrical pulses that can be observed on an
oscilloscope. The depth of the defect (d) of the test object from the surface
is measured by Eq. 6:
where, v is the velocity of ultrasonic wave in the material and t is Time of
Flight (TOF) measurement of the reflected echo. Figure 5 shows
the response of the pristine specimen and the response of the damaged specimen.
The time for the maximum peak of two successive pulses for both specimens was
captured and compared. The presence of defects in the damaged specimen causes
the difference of the duration between two pulses. A 2 kS sec^{1} sampling
rate was used to acquire the signals from the receiver.
EXPERIMENTAL RESULTS
The purpose of this study is to suppress n(t) (Eq. 6) as
the effect of grain noise scattering in ultrasonic testing. The results of the
proposed waveletbased minimization method were compared to the related stateoftheart
methodology, namely SSPAM and SSPPT.
Results from simulated data: Figure 6 shows the outcomes
of the proposed waveletbased minimization algorithm, SSPAM and SSPPT, respectively
on simulated signals. The filter bank in SSP algorithm was made up of ten 100
Hzbandwidth Gaussian filters with 10% overlapping. It can be observed that
waveletbased minimization algorithm outperformed SSPAM and SSPPT to achieve
9.87 dB for SNR. It shows 79.8% improvement of SNR for the simulated noisy data.
It is due to the multiresolutional scaling of the proposed algorithm to highlight
the noise level during signal decomposition. Each iteration of signal decomposition
scales the original data by 2^{j} where, j = 1, 2, 3,..., n. The original
data is in other words zoomed out by the power of two in time domain before
the minimization is carried out.

Fig. 4(ab): 
(a) Noiseless signal with backwall and simulated defect’s
echoes separated by Δt = 4.5 μsec (equivalent to 3.46 mm) and
(b) Noise signal with SNR = 0.1 

Fig. 5(ac): 
(a) GFRP specimen with Mylar induced delamination, (b) Experimental
signal from pristine specimen and (c) Experimental signal from damaged specimen 

Fig. 6(ad): 
(a) Simulated noisy signal, (b) Processed signal by waveletbased
minimization, (c) Processed signal by SSPAM and (d) Processed signal by
SSPPT 

Fig. 7(ab): 
Frequency domain of (a) Experimental signal and (b) Windowed
experimental signal by Hanning window 

Fig. 8(ad): 
(a) Experimental signal from damaged GFRP, (b) Processed
signal by waveletbased minimization, (c) Processed signal by SSPAM and
(d) Processed signal by SSPPT 
This method also works well on separating the closespaced echoes that occurs
especially on thin laminate examination. Figure 6d shows that
SSPPT outperformed SSPAM to obtain higher SNR. SSPPT captures the polarity
reversal for the data in the presence of only grain noise and sets the corresponding
amplitude value to zeroes at time instants. Otherwise, minimum amplitude value
of the data set is depicted. SSPAM is instead based on amplitude averaging
to reduce grain noise.
Results from experimental data: Damage was introduced by inserting a
piece of Mylar between the first and second layer of the GFRP prepreg. The
delamination was approximately 100 mm long and located 0.89 mm depth from the
front surface of the specimen. Ascan data was obtained from the experimental
system discussed. In practical ultrasonic testing, a range of frequency is generated
during pulse excitation. To prevent dispersion that can cause spectral leakage
in frequency spectrum analysis, a Hanning window was applied on the received
data prior to the execution of the proposed algorithm. Figure
7 illustrated the frequency spectrum of the windowed signal. It can be observed
that the frequency width was sharpened. This method separates two frequency
components that are close to each other and have almost equal amplitude to prevent
lose of information in spectrum analysis.
Figure 8 shows the outcomes of the proposed waveletbased
minimization algorithm, SSPAM and SSPPT on experimental data.
Table 1: 
Comparison of the SNR of different methods 

The proposed waveletbased minimization algorithm shows 30.9% improvement of
SNR for the experimental data. The performance of waveletbased minimization
algorithm, SSPAM and SSPPT towards SNR is shown in Table 1.
The proposed method accomplished 9.70 dB for SNR. It can be noted that the SNR
improvement from 7.41 to 9.70 dB for the measured signal is significant. The
proposed method exhibited good performance to detect the defect’s signal
compared to SSPAM and SSPPT.
CONCLUSION
Defect detection in coarsegrain structured material has been a great challenge
in ultrasonic NDE. The backscattered noise due to multicentred pulse reflection
from the grain boundary may hinder the actual defect’s
signal. A multiresolution signal decomposition method based on wavelet principal
has been developed to improve the detection of defect in coarsegrained materials.
This algorithm splits the received signals into multiple frequency bands by
varying the scales of a chosen mother wavelet. This principle works on the hypothesis
that the signal data exhibits polarity reversal when the dataset contains only
grain noise. Thus, for each band of the signal having the same polar direction,
the minimum amplitude value is selected. Otherwise, zero value is chosen. The
processed signals are then recombined to resemble the original signal. The proposed
method has shown to improve SNR in simulated and experimental data. The experimental
results showed that this method successfully resolved the closespaced echoes
in composite materials. This improvement has the benefit of more effective flaw
detection within noise level in complex ultrasound NDE such as composite materials.
As future work, the proposed method is applied on ultrasonic Cscan to enhance
the flaw visibility.
ACKNOWLEDGMENT
This study is supported by the FRGS fund of Malaysia under Grant No. 5523439.