Ultrasonic thickness estimation using multimodal guided lamb waves generated by EMAT

Ultrasonic Thickness Estimation using Multimodal Guided Lamb Waves
generated by EMAT
Joaquín García-Gómez1
, Roberto Gil-Pita1
, Antonio Romero-Camacho2
, Jesús Antonio Jiménez-Garrido2
,
Víctor García-Benavides2
, César Clares-Crespo1
, Miguel Aguilar-Ortega1
1
Signal Theory and Communications Department
University of Alcala
Alcala de Henares, Madrid, Spain
(34) 91-8856751; fax (34) 91-8856699; email roberto.gil@uah.es
2
Innerspec Technologies Europe S.L
Torres de la Alameda, Madrid, Spain
ABSTRACT
The objective of this paper is to study how the selection of the coil and the frequency affects the received modes in
guided Lamb waves, with the objective of analyzing the best configuration for determining the depth of a given
defect in a metallic pipe with the minimum error. Studies of the size of the damages with all the extracted
parameters are then used to propose estimators of the residual thickness, considering amplitude and phase
information in one or several modes. Results demonstrate the suitability of the proposal, improving the estimation of
the residual thickness when two simultaneous modes are used, as well as the range of possibilities that the coil and
frequency selection offers.
Keywords: EMAT sensors, Lamb waves, pipeline inspection, defect sizing, coil selection, frequency selection
INTRODUCTION
Defect sizing in pipeline inspection allows companies to determine when a pipe must be replaced, avoiding costly
repairs in their assets. To tackle this issue, Lamb ultrasonic waves generated through Electro-Magnetic Acoustic
Transducers (EMAT) allow thickness estimation without direct contact with the surface of the metallic material
under investigation [1]. The use of this technology with a meander-line-coil allows generating waves in a directional
way [2], which facilitates differentiating between circumferential and axial scans in Non-Destructive Testing (NDT)
for pipeline inspection [3].
However, the shape of the defect changes the behavior of the ultrasonic signals when they pass through the pipeline,
and it is not easy to predict the amplitude and phase of the wave in function of the residual thickness [4,5]. In recent
studies the use of machine learning techniques applied to information extracted from signals sensed at different
frequencies has been demonstrated to improve the accuracy of the estimation, but the use of multiple frequencies in
general requires more complex sensing devices and more time. A possible way to address these disadvantages is the
use of different modes sensed at a unique frequency, but in this case the selection of the coil and the inspection
frequency becomes a critical aspect, since these different modes must be separable in the measurement, and this is
not always the case.
This paper presents a theoretical study in which the selection of the coil and the frequency for multimodal thickness
estimation are analyzed. The objective is to determine the relationship between the performance of the estimator and
the configuration of the sensing system. The problem was approached from two perspectives. First, a signal
processing based theoretical framework is proposed. Second, simulations obtained by a Finite Element software are
considered. Results demonstrate the suitability of the proposals, improving the estimation of the residual thickness.

LAMB WAVE GENERATION USING EMAT SENSORS
In this section the generation of Lamb waves through EMAT sensors will be described. These sensors are composed
of a magnet and a coil wire. The current is induced in the surface of the ferromagnetic material when the alternating
electrical current flow through the coil wire is placed in a uniform magnetic field near the material. When this field
interacts with the field generated by the magnet, Lorentz force appears. Because of that, a disturbance affects to the
material, creating an elastic wave. If the vibration is coplanar with the propagation plane, these waves are called
Lamb waves. Conversely, the interaction of Lamb waves with a magnetic field induces current in the EMAT
receiver coil circuit.
Lamb waves are characterized by their dispersion and sensitivity to thickness variations. Besides, they can be
divided into modes: symmetric and asymmetric modes. Each mode is composed of two waves (longitudinal and
transversal). They travel at different angles 𝜃" and 𝜃# with velocities 𝑐" and 𝑐#, where the latter refer to the sound
velocity in longitudinal and transversal components, respectively. Considering a Lamb mode that moves in the 𝑥
direction at velocity 𝑐' with a frequency 𝑓, then the wavenumber 𝑘 is related to the longitudinal and transversal
components of the wave:
𝑘" cos 𝜃" = 𝑘# cos 𝜃# = 𝑘 =
2𝜋𝑓
𝑐'
(Eq. 1)
where 𝑘" = 2𝜋𝑓 𝑐" and 𝑘# = 2𝜋𝑓 𝑐# , are the wavenumber of the longitudinal and transversal components,
respectively. Furthermore, the displacement of each wave in the 𝑧 axis can be obtained using 𝛼" and 𝛼#, so that:
𝛼" = 𝑘" sin 𝜃" = 2𝜋𝑓
1
𝑐"
: −
1
𝑐'
:
(Eq. 2)
𝛼# = 𝑘# sin 𝜃# = 2𝜋𝑓
1
𝑐#
: −
1
𝑐'
:
(Eq. 3)
Considering that the wave is reflected in the surfaces and applying the boundary conditions, we can get an equation
related to the dispersion of the Lamb modes. Equation (4) refers to the symmetric modes and equation (5) refers to
asymmetric ones.
4𝑘:
𝛼" 𝛼# sin
𝛼"ℎ
2
cos
𝛼#ℎ
2
+ sin
𝛼#ℎ
2
cos
𝛼"ℎ
2
𝛼#
:
− 𝑘: :
= 0 (Eq. 4)
4𝑘:
𝛼" 𝛼# cos
𝛼"ℎ
2
sin
𝛼#ℎ
2
+ cos
𝛼#ℎ
2
sin
𝛼"ℎ
2
𝛼#
:
− 𝑘: :
= 0 (Eq. 5)
From the previous equations, it can be derived that there exists a relation between the excited frequency 𝑓, the
thickness of the pipe ℎ and the phase velocity 𝑐'. In particular, each mode travels at different 𝑐' depending on the
other above-mentioned parameters. We get a similar relation with the group velocity 𝑐B, defined in equation (6).
𝑐B = 𝑐'
:
𝑐' − 𝑓ℎ
𝜕𝑐'
𝜕𝑓ℎ
DE
(Eq. 6)
Solving the previous equations for different values of frequency and thickness we obtain the phase and group
velocity for each propagating mode.

Now we will consider how signals are generated and received in the pipeline. The EMAT system consists of a
meander-line-coil which generates two signals per loop in the system (one per meander). These waves are
characterized by their wavelength which depends on the separation of the meanders. The following equations are
valid for one mode and then we will iterate for all the modes which appear at a given frequency. Thus, we have to
set the wave equation depending on the group and phase velocities. Considering 𝑓 as the excited frequency, the
transmitted signal propagating in the 𝑥 axis will be generated according to equation (7).
𝑠 𝑥, 𝑡 = sin 2𝜋𝑓 𝑡 −
𝑥
𝑐'
(Eq. 7)
Please note here that the velocity 𝑐' will depend on the frequency and the thickness of the pipe. In a real case, the
transmitted signal includes an envelope 𝑤(𝑡) that generates the transmitted wave packet 𝑝 𝑥, 𝑡 . This envelope
limits the transmission time, and allows controlling the length of the transmitted pulse. Typically, the length of this
envelope is described in function of 𝐶, the number of cycles included in the wave packet. This envelope will travel
at an average velocity of 𝑐B, and in general its shape will change with the distance due to dispersion effects. So, once
the envelope is considered, the transmitted wave packet 𝑝 𝑥, 𝑡 will be expressed using equation (8).
𝑝 𝑥, 𝑡 = sin 2𝜋𝑓 𝑡 −
𝑥
𝑐'
𝑤 𝑡 −
𝑥
𝑐B
(Eq. 8)
From this point, instead of using the transmitted envelope 𝑤(𝑡) we will use using 𝑤(𝑡), which changes its shape in
function of the distance due to dispersion effects. It is also necessary to consider that under EMAT technology the
excitation signal is generated in a set of 𝑁 loops of a coil, separated by a distance 𝐿, which will generate the
propagation wave 𝑦 𝑥, 𝑡 using equation (9).
𝑦 𝑥, 𝑡 = −1 R
sin 2𝜋𝑓 𝑡 −
𝑥 + 𝑚
𝐿
2
𝑐'
𝑤 𝑡 −
𝑥 + 𝑚
𝐿
2
𝑐B
:T
RUE
(Eq. 9)
Each loop generates two signals (one per meander), and the sign of their contribution to the propagation wave
𝑦 𝑥, 𝑡 is included in the term −1 R
. Besides, the measure is sensed at a distance 𝐷, in another set of 𝑁 loops
separated by a distance 𝐿. So, the received signal 𝑧 𝑡 will be expressed using equation (10).
𝑧 𝑡 = −1 RXY
sin 2𝜋𝑓 𝑡 −
𝑥 + 𝑚 + 𝑛
𝐿
2
𝑐'
𝑤 𝑡 −
𝑥 + 𝑚 + 𝑛
𝐿
2
𝑐B
:T
RUE
:T
YUE
(Eq. 10)
The signal received from each mode 𝑧 𝑡 has different values of 𝑐' and 𝑐B, as it was concluded from equations (4),
(5) and (6). Thus, each mode arrives at the receiver with different amplitude and envelope, depending on the
attenuation of each mode and the difference of phase when the signal is received in the coil. Therefore, the amount
of energy of the received signal will vary in function of the frequency.
In order to find out more about the behavior of the modes, a frequency sweep has been carried out between 0 and
800 kHz with one coil and 𝐶 = 4 cycles per wave packet. Figure 1 shows the phase velocity (left) and group
velocity (right), where black color means the energy is maximum at that frequency. Dispersion has been taken into
account to carry out these experiments, since the signal 𝑝 𝑥, 𝑡 has been decomposed with the envelope window

𝑤 𝑡 through the Fourier Transform, and different velocity has been applied to each frequency component. These
graphs correspond to a steel pipe with the following parameters: Young’s modulus 𝐸 = 210 ∙ 10]
𝑁 𝑚:
, Poisson’s
ratio 𝜈 = 0.3 and density 𝜌 = 7800 𝑘𝑔 𝑚a
.
(a) (b)
(c) (d)
(e) (f)
Figure 1: Phase velocity (a) and group velocity (b) in function of the product frequency by thickness, using
coils with different 𝑳 value.

The coil used in the experiments has the following parameters: distance between loops 𝐿 ranging from 0.3 to 0.5
inches, and 𝑁 = 3 loops. It can be observed that the same coil could be used to excite other frequencies, even if it
has been designed to get the maximum energy in a given frequency. Furthermore, if we want to analyze the behavior
of the modes in a deep way, we could change the length L of the coil. In Figure 1 it is observed that the points and
areas of maximum energy vary significantly from one coil to another.
FREQUENCY AND COIL SELECTION FOR MULTIMODAL FEATURE
EXTRACTION
The modeling of the pipeline by means of the ultrasound waves is a non-trivial problem. The changing shape of the
defects makes difficult to draw general conclusions about the relation between the defect and the received signals.
The distortion caused by the defects over the different modes strongly varies with the shape of the mode [4,5]. For
instance, the amplitude of the signal, the time of arrival (group velocity 𝑐B) and the phase velocity 𝑐' of the wrap-
around signal vary with the dimension and shape of the defect.
Thus, it is necessary to analyze how the different modes are going to be represented in the received signal, in order
to look for the best configuration (frequency and size of the coil) that allows a better representation of the different
modes over the same signal.
As it was stated, a meander-line-coil is used to generate the ultrasonic signals that are analyzed once they wrap the
pipeline. It allows us to know the condition of the pipes depending on the different modes and wrap arounds
received. In order to investigate how the behavior of the modes changes according to the length of the coil, a sweep
of experiments has been carried out with coils from 0.30 inches to 0.55 inches, in steps of 0.01 inches. The most
relevant results are shown in Figure 2, where we show where the energy of the different modes is located in a time-
frequency representation. Both asymmetric (A0, A1, A2, A3) and symmetric (S0, S1, S2, S3) modes are plot in
different colors. In each of the modes, curves indicate the area where the energy of the mode is higher or lower.
Results from this figure show that as we change the length of the coil, the parameters related to the modes
(frequency of appearance, area of maximum energy, etc.) are not the same. The final effect is that the modes “move”
in frequency and time. For instance, as the length of the coil in higher, some of the modes appear at lower
frequencies. That is the case of A0, S0, A1 and S1 modes. Other modes disappear from the observed window, such
as the A2 mode (pink), whose second wrap around went away from 0.40 inches to 0.45 inches. First wrap around
disappear from 0.45 inches to 0.50 inches.

(a)
(b)
(c) (d)
(e) (f)
Figure 2: Energy localization of the different modes in function of the frequency.
200 300 400 500 600 700 800 900 1000
Frequency (kHz)
100
150
200
250
300
350
400
450
Time(s)
L=0.30 inches
A0
S0
A1
S1
A2
S2
A3
S3
-20dB
-15dB
-10dB
-5dB
200 300 400 500 600 700 800 900 1000
Frequency (kHz)
100
150
200
250
300
350
400
450
Time(s)
L=0.35 inches
A0
S0
A1
S1
A2
S2
A3
S3
-20dB
-15dB
-10dB
-5dB
200 300 400 500 600 700 800 900 1000
Frequency (kHz)
100
150
200
250
300
350
400
450
Time(s)
L=0.40 inches
A0
S0
A1
S1
A2
S2
A3
S3
-20dB
-15dB
-10dB
-5dB
200 300 400 500 600 700 800 900 1000
Frequency (kHz)
100
150
200
250
300
350
400
450
Time(s)
L=0.45 inches
A0
S0
A1
S1
A2
S2
A3
S3
-20dB
-15dB
-10dB
-5dB
200 300 400 500 600 700 800 900 1000
Frequency (kHz)
100
150
200
250
300
350
400
450
Time(s)
L=0.50 inches
A0
S0
A1
S1
A2
S2
A3
S3
-20dB
-15dB
-10dB
-5dB
200 300 400 500 600 700 800 900 1000
Frequency (kHz)
100
150
200
250
300
350
400
450
Time(s)
L=0.55 inches
A0
S0
A1
S1
A2
S2
A3
S3
-20dB
-15dB
-10dB
-5dB

However, the usefulness of these graphs is that we can set a frequency depending on the modes or wrap arounds we
are interested in, particularly when the objective is not to use just one mode. For instance, if we want to focus on
modes A1 and S1 (dark blue and yellow), it can be seen that 0.30 and 0.35 inches are not the suitable lengths
because both modes will appear mixed in the received signal. We should choose a higher value, such as 0.45 inches,
where first wrap-around of A1 mode as well as first and second wrap around from S1 mode are well separated in
time between 400 and 600 kHz approximately, so they will not be overlapped. Other option would be to choose a
value of 0.50 inches, where these modes are almost completely separated, but including the second wrap-around of
both modes or even the third one from the S1 mode. Again, it is clear that 0.55 inches is not a suitable value because
these modes start to appear together again.
SMART SOUND PROCESSING FOR SIZING ESTIMATION
If we want to solve the problem of pipeline sizing, it is necessary to apply a pattern recognition system, which is
composed of two stages. In the first stage, useful information is extracted from the signals in the form of features.
Later, in a second stage, a predictor tries to learn a model which will be useful for predicting the defects presented in
the pipeline.
To extract useful information from the received signal is very important in the process, since it will be the “raw
material” that the predictor will use. Analyzing a set of signals from real pipelines and the state of the art [6], we
observe that the following features could be useful for the problem at hand:
• Average echo energy (dB), which represents the average energy of the echo received.
• Peak wrap-around energy (dB), which represents the maximum energy of the pulse. We have considered
±30 µs around 𝑡d, the maximum of the signal in the case of absence of defect, to look for the maximum of
each signal.
• Average wrap-around energy (dB), which represents the average energy of the pulse. We have considered
±30 µs around 𝑡d, the maximum of the signal in the case of absence of defect.
• Wrap-around phase delay (µs), which represents how much time has passed between the pulse was sent and
it was received in the same point of the pipeline. It is determined measuring the time difference between
𝑡d and its closest maximum in 𝑧(𝑡). Please note that a delay larger than 1 2𝑓 causes uncertainty, which
conditions the usefulness of this measurement.
• Wrap-around group delay (µs), denoted 𝑡B. In order to estimate this measurement, we consider the centroid
of the average energy of the pulse around 𝑡d, with equation (11).
𝑡B =
𝑡 𝑧(𝑡):efXa∙Edgh
eUefDa∙Edgh
𝑧(𝑡):efXa∙Edgh
eUefDa∙Edgh
(Eq. 11)

Figure 3: Model of the simulated defects.
l
s
d
h

Once we have obtained the features, we need to apply a nonlinear predictor to get the final profile of the pipeline
and to know the performance of the developed model. Neural Networks have been applied, specifically the Multi
Layer Perceptron (MLP) [7]. In this paper MLPs with a hidden layer of twenty neurons have been trained using the
Levenberg-Marquardt algorithm [8].
From the results presented in Figure 2, we will select and 𝑓 = 450 and a coil with 𝐿 = 0.52 inches. So, in the case
of using just the main mode we will consider 5 features, and in the case of considering two modes we will have 9
features (4 wrap-around features for each mode plus the echo energy).
To study the relationship between these parameters and the shape of the defects, we have used the Finite Element
Method (FEM) included in the Partial Differential Equations Toolbox of Matlab. With these simulations we have
generated a database with several different defects. The defects have been characterized with three parameters:
length (𝑙), depth (𝑑) and slope (𝑠). Figure 3 describes the meaning of these parameters in a real pipeline. The
thickness of the pipe used is ℎ = 7.8 mm, and the distance to the receiver is 𝐷 = 0.7 m. For simplicity, we have not
modeled the width of the defect, that is to say, we have not considered the 𝑦 dimension of the pipeline.
Table 1: RMSE (mm) in the estimation of the residual thickness using an MLP with 20 neurons in the hidden
layer for different number of used modes.
S1 mode
5 features
S1 and A1 modes
9 features
RMSE (mm) 10.50 mm 9.72 mm
In a second approach, we have developed an experiment using a synthetic database for estimating the residual
thickness of the pipeline. The database consists of 384 signals generated with defects of different shape. The length
of the defect (𝑙) ranged from 10 to 100 mm, the depth (𝑑) from 0 to 9 mm, and the slope (𝑠) from 1 to 100 mm.
To obtain the prediction results, 𝑘-fold cross validation was applied in the generated database, being 𝑘 = 5. This
method consists in dividing the database in 𝑘 groups so that the full process is repeated 𝑘 times, using one group of
signals as test subset and the remaining 𝑘 − 1 groups as training subset. Results are then averaged to obtain the Root
Mean Square Error (RMSE) of the estimation of the depth. The advantage of this method is that the obtained results
are generalizable to defects different from those used in the database.
The objective is to know how well estimated is the received signal at different frequencies, so different experiments
have been considered. First, we have considered the use of only one frequency, and we have also studied what
happens when both frequencies are used at the same time. Concerning the features, the usefulness of each feature
has been studied, and the inclusion of all three features has also been considered. Table 1 shows the RMSE in
function of the features and the frequencies. In all the cases above it is clear that as we get better results when two
modes are used.
CONCLUSIONS
Pipeline inspection problem can be approached in many different ways. Lamb wave generation through EMAT
sensors proves to be a very effective and useful one. However, the amount of information provided by the wrap-
around signals needs to be processed by advanced techniques, such as smart sound processing algorithms. Thanks to
them, it is feasible to get good estimation results of the pipeline defects, in both real and simulated signals.

In this paper we establish tools for determining the frequency and dimensions of the coil in order to be able to
analyze two modes with an unique scan. We study how the behavior of the modes change when the length of the
used coil is different, demonstrating its interest for multimodal approaches. Studies of the size of the damages with
all the extracted parameters have been used to propose estimators of the residual thickness, considering amplitude
and phase information. Results with two modes demonstrate the suitability of the proposal, improving the estimation
of the residual thickness.
ACKNOWLEDGEMENTS
This work has been funded by Innerspec Technologies Europe S.L through the “Chair of modeling and processing
of ultrasonic signals” (CATEDRA2007-001), and by the Spanish Ministry of Economy and Competitiveness-
FEDER under Project TEC2015- 67387-C4-4-R.
CONFLICTS OF INTEREST
The authors declare that there is no conflict of interest
REFERENCES
(1) Green, R.E., 2004. “Non-contact ultrasonic techniques”. Ultrasonics, 42, 9–16.
(2) Zhai, G., Jiang, T., Kang, L., 2014. “Analysis of multiple wavelengths of Lamb waves generated by meander-
line coil EMATs”. Ultrasonics, 54, 632–636.
(3) Salzburger, H.J., Niese, F., Dobmann, G., 2012. “EMAT pipe inspection with guided waves”. Welding in the
world, 56, 35–43.
(4) Demma, A., 2003. The interaction of guided waves with discontinuities in structures. PhD thesis, University of
London.
(5) Cobb, A.C., Fisher, J.L., 2016. “Flaw depth sizing using guided waves”. AIP Conference Proceedings. AIP
Publishing, Vol. 1706, p. 030013.
(6) García-Gómez, J., Bautista-Durán, M., Gil-Pita, R., Romero-Camacho, A., Jimenez-Garrido, J.A., Garcia-
Benavides,V., 2018, “Smart Sound Processing for Residual Thickness Estimation using Guided Lamb Waves
generated by EMAT”. 27th
ASNT Research Symposium, 99-105.
(7) Weisz, L., 2016. “Pattern Recognition Statistical Structural And Neural Approaches”. Pattern Recognition, 1,
2.
(8) Hagan, M.T.; Menhaj, M.B., 1994. “Training feedforward networks with the Marquardt algorithm”. IEEE
transactions on Neural Networks, 5, 989–993.

Ultrasonic thickness estimation using multimodal guided lamb waves generated by EMAT

More Related Content

What's hot (18)

Similar to Ultrasonic thickness estimation using multimodal guided lamb waves generated by EMAT (20)

More from Innerspec Technologies (20)

Recently uploaded (20)

Ultrasonic thickness estimation using multimodal guided lamb waves generated by EMAT