Ultrasound material backscattered noise analysis by a

duo wavelet-regression analysis

F. Bettayeb
Research center on welding and control, CSC, Route de Dely Brahim, Bp:64, Cheraga, 16800
Algiers, Algeria
fairouz [email protected]

2857
Internal material defects detection by ultrasound non destructive testing is widely used in industry, ultrasonic
data are obtained from travelling waves inside the matter and captured by piezoelectric sensors. The natural
inhomogeneous and anisotropy character of steel made material causes high acoustic attenuation and scattering
effect. This adds complexity to data analysis. In this research we address the non linear features of back scattered
ultrasonic waves from steel plates and welds. Indeed structural noise data files captured from specimens, and
processed by a wavelet energy filtering approach, show significant insights into the relationship between
backscattered noise and material microstructures. This algorithm along with correlation coefficients, residuals
and interpolations calculations of processed ultrasonic data seems to be a well-adapted signal analysis tool for
viewing material micro structural dimension scales. Experiments show a challenging 3D interface between
material properties, calculations and ultrasonic wave propagation modelling. As well as they indicate a quasi
linear signal energy distribution at micro structural levels. It suggests probable incidence of microstructure
acoustic signatures at different energy scales of the material phases. Multi polynomial interpolations of the
processed noise data exhibit an attractor shape which should involves chaos theory noise data.
1 Introduction
Acoustical characterization is an important item in
materials testing; it takes significant status during
fabrication and in service inspection process. Ultrasonic
techniques have been commonly used in power and
petrochemical industries for nearly 50 years. However, cast
or welded austenitic components remain difficult to reliably
and effectively examine. In some devices grains
orientations produce ultrasonic beam divergence and
splitting mainly in the case of multi-pass welds when the
re-melting process after each pass causes complex
solidification process. Anisotropic grains large size
compared with acoustic pulse wavelength, affects coarsely
ultrasound propagation; by causing severe attenuation,
changes in velocity and energy scattering [1]. Sound beam
refraction and reflection arising at grain boundaries induce
defects incorrectly reported, specific volumes of materials
not examined or both [2]. Various industrial inspections on
dissimilar components confirm the consequence of these
physical phenomena on ultrasonic inspection
implementation. Some experimental studies as in [3] and
[4] confirm grain size influence on attenuation and noise, in
addition to a frequency filtering when the wavelength is
equal to the average grain diameter. Therefore, it seems to
be essential to make relationships between material micro
structural features and ultrasonic beam acoustic
characteristics. And try to examine micro structural
parameters which are the source of attenuation and
structural noise origin. As ageing and environment
consequence on failure mechanisms cannot be sufficiently
predicted by traditional methods, computational modelling
of materials behaviour is becoming a reliable tool to
emphasize scientific investigations and to match up
theoretical and experimental approaches. This requires not
only development of improved processing techniques but
also better understanding of material structure. These
conditions implicate multiple length scales analysis and
multiple implementation steps.
In this paper we present a new structural noise features
analysis based on an energy smoothing algorithm. The new
de-noising algorithm performs an accurate signal analysis
as well as detection of little defects of 1mm. The following
experiments obtained from structural noise signal captured
from a steel plate, will give significant insights into the
relationship of backscattered noise and microstructures
which can help to micro structural dimension scales
understanding
2 Ultrasonic s noise features :
overview and challenge
Since ultrasonic signal is transient, non-stationary, and
limited in time and frequency, extraction and analysis of the
useful information remain difficult. Basically, flaw
visibility is corrupted by electrical, pulse, ringing, structure
noises or spurious signals. Commonly acoustic noise is
assumed to be gauss random variable with zero averaging
and limited band power spectrum function [1] [2]. Various
signal processing techniques were investigated to interpret
waveform data and extract useful information for further
diagnostic and predictive purpose. In the literature, there
are three main categories of waveform data analysis: time
domain, frequency domain and time frequency analyses.
The first calculates typical attributes as descriptive statistics
(mean, peak, standard deviation, high order statistics etc.),
and extract features by the use of autoregressive parametric
models. However the complexity of the model order
estimation carries on complicated modelling. In frequency
domain, spectral analysis is certainly the oldest technique
that presents the hidden view of the signal. Incompatible for
transitory signals and non stationary data, its efficiency is
limited. To solve this problem, time-frequency distribution
approved several reliable techniques such as short-time
Fourier transform (STFT), Wigner-Ville distribution and
wavelet transform. Wavelet transform is one of the most
successful processing techniques able to withdraw the non
stable characteristics of the signal.
Several applications of wavelet transform for defect
detection were proposed, using continues wavelet transform
enriched with recent techniques, discrete or multiresolution
analysis and wavelet packet transform [9] [10]. An
interesting synthesis of these techniques is presented in
[11]. Similar to the time frequency distribution, wavelet is a
time scale representation. It expresses the signal in a series
of oscillatory functions with different frequencies at
different times. Its main advantage is its ability to produce
high frequency resolution at low frequencies and high time
resolution at high frequency, for signal with long duration
low frequencies and short duration high frequencies. This
provides facilities to noise cancellation in natural signals
[2] [6] [7].
The aim of this work is to propose a new method based
on wavelet analysis optimization. In this paper, wavelet
multiscale analysis was investigated with a forecast
viewpoint, as a powerful computational tool for noise
discrimination and features extraction. This idea has

2858
emerged after having examined continuous, discrete,
wavelet packet and dual transforms on natural signals from
heterogeneous materials, mostly welding defects signals.
More details of this works are given in [12] [13] [14]. In
this work the structural noise features are extracted by a
new energetic smoothing algorithm which permits the
identification of the noise analyzing function and
invalidation of the noise random nature. The energetic
extraction of the noise and the useful signal has provided
easy filtering with enhanced defect detection in natural
ultrasonic signals from steel pieces with artificial flaws and
welding defects
3 The energy approach
Since useful ultrasonic energies are clustered in the
signal central frequency band and the defects energy fit
lower frequencies than the structural noise. So the energetic
analysis provides a larger view of the signal energetic
configuration and permits as well as easy extraction of the
noise.
Wavelet basis functions as described in the Mallats
book [7] are constructed by dyadic dilation (index j) and
translation (index k) of a mother wavelet:

(1)

Wavelet transform is characterized by two functions the
scaling function (2) and its associated wavelet (3):

(2)


k)
(3)

Where g (k) is a suitable weighting sequence and h (k)
is the refinement filter
The orthogonal aspect of wavelet transform provides for
any function f(x) vectors of atoms composed by detail
coefficients (4) and approximation coefficient (5) which
characterize the atomic decomposition of f(x):

(4)
(5)

Subsequently wavelet transform with a depth j can be
adjusted by (6):


jk
(6)

3.1 From atomic representation to energy
distributions

The purpose of the energy distributions is to distribute
the energy of the signal over time and frequency. The
starting point is that since the energy of a signal x can be
deduced from the squared modulus of either the signal or its
Fourier transform, we can interpret |x(t)|
2
and |x(v)|
2
as
energy densities, respectively in time and in frequency. It is
then natural to look for a joint time and frequency energy
density
z
(t, v), such that:


(7)

3.2 Wavelet smoothing method
Smoothing is an estimation technique that takes into
account both past and future observations, and can be more
accurate than filtering [22]. Generally, wavelet smoothing
and wavelet de-noising can be used to enhance signal to
noise ratio. The difference between these two processes is
that smoothing removes high frequency components of the
transformed signal regardless of their amplitudes, while
denoising removes small amplitude components of the
transformed signal regardless of their frequencies.
However, it is not easy to choose a suitable strand value for
de-noising which is significant to the noise suppression
achievement without signal loss. For this reason, wavelet
smoothing provides fine visual quality of the processed
scales (spectra), which is more suitable to signal features
extractions. Sachs in [19] gives a rich report on wavelet
smoothing by non linear thresholding for non stationary
time series de-noising and signal recovery. In any case, the
literature designs two classes of smoothers: linear,
including local polynomial smoothing, loess, spline and
kriging, and nonlinear, such as running medians and other
median-based smoothers [20] [21] [22]. In contrast to their
performance for data containing only Gaussian noise, linear
smoothers do not respond well to data containing impulsive
noise, or noise generated by microstructures. Non linear
energetic smoothing algorithms are more suitable.
4 New approach and algorithm
If the above methods are suitable their implementation
needs several algorithms and experiments, for detecting
best analysing functions and best threshold regulation rules.
In fact Hwang, Mallat theorem indicates the presence of
maxima at the finer scales where singularities occur, in
addition when the wavelet is the n
th
derivative of a
Gaussian, the maxima curves are connected and go through
all of the finer scales [7]. As the 8
th
Gaussian derivative is
the analysing function in our experiments, the core of the
Hwang, Mallat theorem offers us the opportunity to
investigate the spirit of the minima maxima smoothing
energetic analysis.
The new filtering algorithm is based on the energy
content of the wavelet coefficients via an energy smoothing
of the noise function [23].

4.1 The algorithm

While ultrasonic energies are concentrated in the central
frequency band, therefore different frequencies close the
band are represented in the transform domain by very weak
amplitudes and can be scattered without loss of
information.
But how the structural noise can be removed? The idea
is to approximate it with an analysing function. The
proposed algorithm illustrated admits the development of a
noise analysing function with an easy filtering process. In
this algorithm, the extraction from the signal of the noise
energetic coefficients is based on the removal of the
maximum energetic coefficients vector from the original
signal wavelet decomposition by the 8
th
derivative gauss

2859
function. In contrast the computation of the noise energetic
threshold is achieved from the wavelet coefficients of the
noise Morlet scalogram. An inverse wavelet transform
procedure gives us statistical noise characterization. The
Morlet function is selected after a correlation process
between wavelet bases and extracted noise database from
ultrasonic signals captured from welds, welding defects and
artificial flaws. Then the filtering is performed based on an
energetic subtraction of the maximum noise energetic
coefficients vector analysed by the Morlet, from the
minimum signal energetic coefficients vector analysed by
the 8
th
derivative of the Gaussian i.e. a subtraction between
two continuous wavelet representations of the same signal
is performed.
5 Experiments
Steel material used in these experiments is a rich
element which can undergo quenching and tempering, see
chemical analyses in table 1. Metallographic investigations
reveal ferrite and pearlite structure (figure 1). Grain size
varies between 40 and 60 m and hardness testing gives an
average value of 120 HV. Structure noise function extracted
and analyzed by the de-noising algorithm [9], undergoes
computing mismatching and correlation process between
interpolation and residuals coefficients. See example of
pure signal filtering in (Figure 2). Obtained results, point to
occurrence of quasi linear energy distribution in (figure 3),
which could advise to apparent energy scales incidence of
micro structural acoustic signatures. In (figure 4) residuals
display fitting indications from first samples noise data
obtained after multi-interpolation stages. This must be
correlated with ultrasonic frequency band and material
behavior. (Figure 5) reveal particular residual distributions
of Fourier transform noise function, after several
polynomial interpolations. This will recommend
relationship exploration with some material properties.



Figure 1. Ferrite pearlite structure (X50 & X100)

(a)
0 200 400 600 800 1000 1200
-6
-4
-2
0
2
4
6

(b)
0 200 400 600 800 1000 1200
-60
-40
-20
0
20
40
60
bruit


sr

(c)
0 200 400 600 800 1000 1200
-400
-300
-200
-100
0
100
200
300
400


Figure 2. Signal and noise analysis: (a) Natural signal from
steel piece of 35 mm, inside 1 mm flaw indications
captured by a 5 MHz Krautkramer transducer. (b)
Extracted functional noise. (c) Smooth filtered signal where
the flaw indications are amplified and noise totally
withdrawn.

200
400
600
800
1000
5
10
15
20
25
-0.8
-0.6
-0.4
-0.2
0
0.2
0.4
0.6


-0.6
-0.4
-0.2
0
0.2
0.4
0.6


Figure 3. Noise functional analysis displays different
energy scales

10
-1
10
0
10
1
10
2
10
3
-150
-130
-110
-90
-70
-50
-30
-10
10
30
50
70
90
110
130
150
1024 samples of the noise signal (log)
residuals


Spline: normof residuals = 0
4thdegree: normof residuals = 284.1592
5thdegree: normof residuals = 284.1442
6thdegree: normof residuals = 284.1176
7thdegree: normof residuals = 284.0955
8thdegree: normof residuals = 284.0493
9thdegree: normof residuals = 284.0013
10th degree: normof residuals = 283.9499
9th degree
10th degree
7th degree
5th degree
8th degree
6th degree
4th degree
spline

Figure 4. Residuals of the structure noise after several
interpolations

-60 -40 -20 0 20 40 60 80 100 120
-80
-60
-40
-20
0
20
40
60
80


signalfft
quadratic
cubic
4th degree
5th degree
6th degree
7th degree

Figure 5. Attractor shape after multi-polynomial
interpolations

Table 1: XRF Chemical analysis

Cr 1.36%
Ni 0.90%
Mn 0.85%
Si 0.52%
Al
Cu
0.45%
0.13%
S 0.00218%
Mo 0.048%

2860
Mg 0.04%
Sb 0.02%
P 0.01%
V 0.01%
Fe balance
C 0.126%
6 Conclusion
Non linear denoising of ultrasonic signals captured from
welds, with multiscale approximation using thresholding,
permits an adaptive representation of the signal
discontinuities. The new energy algorithm involving the
energetic matter of the signal and the noise, by means of
minimisation of a smoothing functional is promising. In this
algorithm no signal decomposition is performed and the
threshold level is determined by an arithmetic process of
the maximum and the minimum wavelet coefficients
energetic level. Therefore the structural noise is
approximated by a wavelet function, and the denoising
process is carry out by discrimination between two wavelet
functions. This algorithm is powerful when the selected
analyzing functions are the best matching mother wavelet
functions to signal and noise information. In reverse case, a
scaling function must be composed for the generation of the
experimental wavelet functions. The approximation of the
structure noise by the Morlet function, offers the prospect
to investigate a multiscale material microstructure
characterization, in an attempt to extract some useful
microstructure material features as presented in Fig 9,
where we can observe different levels of the structural noise
energy concentration at different scales, extracted from a
steel plate by ultrasonic testing with 5Mhz piezoelectric
tranducer. In reality, if anisotropic noise is related to local
variations in texture or shapes of macro etches, the
relationship of this ultrasonic property to microstructure is
not well understood, and up to now no careful theory has
been presented to quantitatively describe these relationships
[24]. Chaos theory seems to be helpful for this issue.
References
[1] Mike C. Tsao, Measurement and analysis of ultrasonic
beam profiles in a solid. Material evaluation, 1998;
Vol. 56, N 5
[2] Abbate, J. Koay, J. Frankel, Signal and noise
suppression using a wavelet transform signal
processor: application to ultrasonic flaw detection.
IEEE, Trans. ultrasonic, ferroelectrics, and frequency
control, 1997; Vol. 1, N 1
[3] Schumm, A., Chassignole, B., Diaz, J., Duwig, V.,
Fouquet, T., Structural Noise in modelisation, ECNDT,
2006
[4] Jenson, F., Poidevin, C., Doudet, L., Modlisation du
contrle ultrasonore dans une structure gros grains,
Journes COFREND, 2008
[5] Mallat, A theory for multi-resolution signal
decomposition: the wavelet representation. IEEE
Trans. pattern analysis and machine intelligence, 1989;
Vol.11, N7:674-693
[6] C.S. Burrus, R. A. Gopinath, H. Guo, "Wavelets and
wavelet transforms", Rice University, 98; Houston
Edition.
[7] S. Mallat, A wavelet tour of signal processing, 1999;
second edition, academic press, New York.
[8] P.C.Ching, H.C. So, S.Q. Wu; Wavelet de-noising and
its application to time delay estimation. IEEE Trans.
signal processing, 1999; Vol. 47, N10
[9] Rubini, R., Meneghetti, U. Application of the envelope
and wavelet transform analyses for the diagnosis of
incipient faults in ball bearings.-Mechanical Systems
and Signal Processing, 2001, 15(2), p.287-302.
[10] Ericsson, E. and al. Towards automatic detection of
local bearing defects in rotating machines.-Mechanical
Systems and Signal Processing, 2005, 19, p.509-535.
A.K.S.
[11] Jardine, D. Lin, D. Banjevic, A review on machinery
diagnostics and prognostics implementing condition
based maintenance, Mechanical system and signal
processing, 2006, vol. 20 n6,p.1483-1510
[12] F.Bettayeb, S. Aoudia, S. Haciane, Improving the time
resolution and signal noise ratio of ultrasonic testing of
welds by the wavelet packet, NDT&E international,
2005; Vol. 38: 478-484, Elsevier
[13] F.Bettayeb, T. Rachedi, H. Benbartaoui, An improved
automated ultrasonic NDE system by wavelet and
neurone networks, Ultrasonics, 2004; Vol.42: 853-
858, Elsevier
[14] F. Bettayeb, D. Benbachir, K. Boussiha, An energetic
smoothing analysis for the ultrasonic signal de-noising
and defect detection, in proceeding of the 9th ECNDT,
2006; BB 103-CD, ISBN 3-931381-86-2, published &
copyright by DGZfp e.V
[15] A.R. Ferreira Da Silva, Wavelet denoising with
evolutionary algorithms, Digital signal processing,
2005; Vol.15, issue 4: 382-399, Elsevier
[16] E. Pardo, J.L. San Emeterio, M.A. Rodriguez, A.
Ramos, Noise reduction in ultrasonic NDT using
undecimated wavelet transforms, Ultrasonics,2006;
Vol. 44:1063-1067, Elsevier.
[17] Sung, C.K., Tai, H.M. and Chen, C.W. Locating
defects of gear system by the technique of wavelet
transform. Mechanism and machine theory, 2000, 35,
p.1169-1182
[18] Donoho, D. L, De-noising by soft-thresholding, IEEE
Transactions on Information Theory, 1995; Vol. 41, N
3:613-627
[19] R.V. Sachs, K. Schneider. Wavelet Smoothing of
Evolutionary Spectra by Nonlinear Thresholding.
Applied and computational harmonic analysis, 1996, 3,
p. 268-282, Academic Press, Inc.
[20] Bowman, A.W. & Azzalini, A. (1997). Applied
Smoothing Techniques for Data Analysis, Oxford
University Press, London.
[21] Daubechies. Ten Lectures on Wavelets. SIAM, 1992.
[22] Donoho, D.L. & Johnston, I.M. (1995). Adapting to
unknown smoothness via wavelet shrinkage, Journal of
the American Statistical Association 90, 12001224.
[23] F. Bettayeb, Wavelet threshold enhancement by an
energetic ultrasonic signal characterization, in
proceeding Acoustics08 Paris, 2008 p: 2057-2060
[24] Y.K. Han, R.B. Thomson, ultrasonic backscattering in
duplex microstructures: theory and application to Ti
alloys, Metallurgical and materials transactions,
Vol.28A, Jan 1997, pp: 91-104.
[25] Rohwer. Nonlinear Multiresolution Analysis. 2004,
Birkhaser Publishers.


2861