1958 IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, VOL. 57, NO.

9, SEPTEMBER 2008

Crack Shape Reconstruction in Eddy Current Testing

Using Machine Learning Systems for Regression
Andrea Bernieri, Luigi Ferrigno, Member, IEEE, Marco Laracca, Member, IEEE, and Mario Molinara

Abstract—Nondestructive testing techniques for the diagnosis of allows the shape and geometrical dimensions of the defect
defects in solid materials can follow three steps, i.e., detection, loca- to be evaluated with acceptable uncertainty. On this basis, an
tion, and characterization. The solutions currently on the market acceptance/rejection decision is carried out by either a human
allow for good detection and location of defects, but their char-
acterization in terms of the exact determination of defect shape operator or an automatic system.
and dimensions is still an open question. This paper proposes a Defect detection and location are the state of the art in this
method for the reliable estimation of crack shape and dimensions field, but the exact determination of the shape and dimensions
in conductive materials using a suitable nondestructive instrument of the fault area is still in the research domain, as confirmed by
based on the eddy current principle and machine learning system the growing attention of the international scientific community
postprocessing. After the design and tuning stages, a performance
comparison between the two machine learning systems [artificial [6]–[12]. Typically, in the ECT field, component selection is
neural network (ANN) and support vector machine (SVM)] was carried out by expert human operators that decide the accept-
carried out. An experimental validation carried out on a number ability of the piece under test by evaluating “rule of thumb” use-
of specimens with different known cracks confirmed the suitability ful quantities, such as typical 2-D graphs that show the behavior
of the proposed approach for defect characterization. of reactance versus the resistance of the impedance response
Index Terms—Artificial neural network (ANN), eddy current of the ECT probe (the measured impedance plane). Modern
testing (ECT), nondestructive testing (NDT), signal processing, instruments are able to automatically select the bad piece using,
support vector machine (SVM). for example, a comparator, which analyzes the measured signal
with reference to a suitable threshold. More sophisticated and
I. I NTRODUCTION expensive instruments are based on digital technology: They
can carry out tests and remotely acquire data, run sophisticated
N ONDESTRUCTIVE testing (NDT) is a very broad inter-
disciplinary field that plays a critical role in assuring
that structural components and systems perform their function
analysis procedures, control robotics, sound alarms about cer-
tain conditions, and provide an output to mark, cut, or reject a
defective part. These instruments also provide information on
in a reliable and cost-effective way. As requested by product
particular kinds of defects but, in any case, neither reconstruct
norms [1], NDT technicians and engineers define and imple-
crack shape nor evaluate its geometrical characteristics (the
ment tests that locate and characterize material conditions and
characterization step) [13]. This information is very important
flaws. These tests are performed in a manner that does not
in deciding the acceptability of pieces under test and finding the
affect the future usefulness of the object or material. Common
origin of defects both for economic and safety reasons.
technologies for NDT equipment include radiography or X-ray
Different solutions have been proposed to solve this problem
analysis, laser holography or gaging, penetrant testing, mag-
[7]–[12]. Most of them are centered on model-based methods,
netic particle testing, and eddy current testing (ECT).
in which a suitable electromagnetic model is used to reconstruct
Recent advances have made ECT more powerful, versatile,
crack shapes. The other solutions are based on model-free
and useful for quality assurance. In fact, modern ECT tech-
methods, which use suitable digital signal processing (DSP)
niques offer unique low-cost methods for the high-speed large-
(such as ECT camera systems, image analysis methods, neural
scale inspection of metallic materials, such as those used in
network applications, and so on) on acquired data obtained
nuclear, aerospace, and maritime applications, where premature
from the material under test.
failures could give rise to economic problems and/or endanger
Model-based methods (usually based on the finite-element
human life [2]–[5].
method) manifest a number of problems related to high process-
The NDT eddy current (EC) tests can be divided into three
ing complexity and long processing times, and their use in
steps, i.e., detection, location, and characterization. Detection
online applications is difficult. Model-free methods, on the
is the ability to find the presence of a defect; location finds a
other hand, show low complexity and short processing times
specimen area where the defect is situated; and characterization
but should be accurately tuned for the particular application.
Taking these considerations into account, the authors have
Manuscript received July 15, 2006; revised February 5, 2008. already proposed a model-free method for the reliable estima-
The authors are with the Department of Automation, Electromag-
netism, Information Engineering, and Industrial Mathematics, University of tion of crack shape and dimensions based on the integration of
Cassino, 03043 Cassino, Italy (e-mail: [email protected]; [email protected]; an EC NDT instrument and a support vector machine (SVM)
[email protected]; [email protected]). processing algorithm [14]. This method shows good general-
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org. ization capability, acceptable performance in defect evaluation,
Digital Object Identifier 10.1109/TIM.2008.919011 and good processing time.

0018-9456/$25.00 © 2008 IEEE

Authorized licensed use limited to: Centrale Supelec. Downloaded on October 08,2024 at 10:37:32 UTC from IEEE Xplore. Restrictions apply.
BERNIERI et al.: CRACK SHAPE RECONSTRUCTION IN ECT USING MACHINE LEARNING SYSTEMS 1959

field is sensed by the EC probe, whereas built-in DSP-based

software highlights the magnetic field variations due to the pres-
ence of a crack. Since the presence of defects causes asymmetry
in the induced signals, processing software in the frequency
domain can highlight these variations [15]. In particular, by
means of a suitable calibration step, it is possible to relate the
magnetic field values to the estimated harmonics [16].
The amplitudes and phases of the resultant magnetic field
are stored in the NDT instrument memory, together with their
geometrical coordinates. The aforementioned steps are repeated
until the whole specimen area is scanned. Suitable magnetic
field maps are then obtained. To exemplify the procedure
used, we consider the aluminum specimen, the defect, and the
scanning path described in Fig. 2. Two maps describing the
experimental magnetic field amplitude of the sensors oriented,
respectively, along the x-axis [Fig. 3(a)] and y-axis [Fig. 3(b)]
are then produced. For the sake of brevity, the magnetic field
phase maps are not provided here since examples can be found
in [15].
The total amount of data to form suitable magnetic field
maps is then passed on to the processing section. This extracts
information features and performs defect characterizations by
means of a proper measurement algorithm based on a machine
learning system.
The following subsections provide details regarding the bi-
axial EC probe, the scanning system, the NDT measurement
instrument, and theoretical remarks on both the ANN and SVM
software approaches.

A. Biaxial EC Probe
The biaxial EC probe comprises two sensors that are able to
Fig. 1. Block diagram of the proposed measurement method. detect magnetic fields on plane regions with good sensitivity
(5 µV/µT). The core of each sensor is the fluxset [17], [18],
This paper presents further improvement on the above which is a device that comprises two coils (driving and pickup)
method in terms of both measurement performance and gen- coaxially wired onto a thin ferromagnetic sheet of annealed
eralization capability. The results of the proposed method are metallic glass [a soft magnetic material with high initial per-
compared with those furnished by an alternative learning ma- meability (µr
85 000) and a low saturation field (0.65 T)]. In
chine, i.e., an artificial neural network (ANN), in terms of the absence of an external field, the excitation of the driving coil
measurement accuracy, generalization capability, robustness, induces a symmetric voltage waveform on the pickup coil. This
and computational burden. waveform becomes increasingly asymmetric as the magnetic
A description of the proposed method (the SVM and ANN field measured by the sensor increases. By adding an additional
processing algorithm adopted) and the tuning stage in a simula- (excitation) coil supplied by a sinusoidal current to induce an
tion environment is detailed in the following sections. A num- EC in the specimen, a complete NDT EC probe can be obtained.
ber of experiments were performed on specimens with known
cracks to verify the performance of the proposed method. B. Scanning System
The adopted scanning system uses two motorized axes that
II. P ROPOSED M EASUREMENT M ETHOD allow a range of motion equal to 300 mm × 300 mm in the
x−y plane with a resolution of 0.1 mm. The repeatability is
A block diagram of the proposed measurement method is ±0.03 mm for each axis with a backslash of 0.03 mm. A
provided in Fig. 1. Two of its sections can be highlighted here, micrometer jack screw allows accurate positioning in terms
i.e., the measurement and processing ones. The measurement of vertical distance (z plane). Software driver libraries allow
section comprises a suitable NDT instrument connected to a adaptive scanning strategies to be set up.
precise scanning system equipped with a biaxial fluxset-based
EC probe.
C. NDT Measurement Instrument
The scanned specimen area and the step resolution are fixed;
the probe is moved on the specimen, and for each position, ECs As mentioned above, the adopted NDT instrument, which
are induced in the specimen under test. The resultant magnetic was already proposed in [19], is able to drive the scanning
Authorized licensed use limited to: Centrale Supelec. Downloaded on October 08,2024 at 10:37:32 UTC from IEEE Xplore. Restrictions apply.
1960 IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, VOL. 57, NO. 9, SEPTEMBER 2008

Fig. 2. Aluminum specimen with the path, crack position (x0 , y0 , z0 ), and its geometrical characteristics (length l and height a).

Fig. 3. Measured magnetic field amplitude maps of (a) the x-axis sensor and (b) the y-axis sensor. The x- and y-axes in the figures are the scan directions.

system to induce ECs in the specimen and to detect the reaction converter channels with a maximum sampling rate of 20 MHz.
magnetic field. Its modular architecture comprises five sections, The above-described DSP-based architecture processes the ac-
which are outlined in block diagram in Fig. 4 and can be quired signals to compute the reaction magnetic field.
described as follows: An important feature of the instrument setup is its built-
in self-adjustment capability, which allows the magnetic noise
1) the waveform generation unit (WGU), to excite driving
and direct current fields present in industrial environments to
coils on the EC probe;
be dynamically compensated, thereby rendering the instrument
2) the excitation coil unit (ECU), to supply the excitation
suitably calibrated for different measurement conditions.
coil to induce the EC in the specimen;
3) the motor driver unit (MDU), to move the EC probe;
4) the acquisition, elaboration, and control unit (AECU), D. Processing Algorithm
which will be described in more detail hereinafter;
The capabilities of two machine-learning-based methods,
5) the power supply unit (PSU), to feed all the previously
such as ANN and SVM, were investigated. Both methods can
mentioned units.
solve i) classification and ii) regression problems by learning
The AECU unit is the core of the instrument, since it acts through the examples used in training. It should be remarked
as a system controller and manages all the measurement and here that i) classification separates a large class of objects into
communication functions during the measurement process. It is smaller ones and gives a criterion for determining whether
based on a floating-point DSP with an operating frequency of a particular object is or is not in a particular class, whereas
40 MHz, a 256-Kword RAM, and three 12-bit analog-to-digital ii) regression predicts the value of a continuous output variable.
Authorized licensed use limited to: Centrale Supelec. Downloaded on October 08,2024 at 10:37:32 UTC from IEEE Xplore. Restrictions apply.
BERNIERI et al.: CRACK SHAPE RECONSTRUCTION IN ECT USING MACHINE LEARNING SYSTEMS 1961

is negligible compared to the others. 2) The crack trajectory

is linear. 3) It has a rectangular shape (with a1 = a2 = a and
l1 = l2 = l), positioned parallel to the surface (z1 = z2 = z).
4) It can be superficial (z0 = 0) or subsuperficial (z0 = 0),
where the coordinate z0 is the crack depth in relation to the
top surface of the specimen.
Using this assumption, the crack can be completely described
using five parameters, i.e., the x, y, and z geometrical co-
ordinates (x0 , y0 , z0 ) of its starting point and its length (l)
and height (a). These five continuous variables have to be
accurately estimated and can be considered as goals of the mea-
surement algorithms. A regression problem, therefore, needs to
be solved.
Since an ANN is naturally defined as a regression method,
which is well described in the literature [20], [21], its general
description is not reported in this paper. On the contrary, an
SVM is naturally designed to solve classification problems, and
its use in regression is quite recent. A description of an SVR,
which is an SVM for regression, is, therefore, reported in the
Appendix.

Fig. 4. Architecture of the proposed NDT instrument, where PSU is the power
supply unit, ECU is the excitation coil unit, WGU is the waveform generation E. Processing Algorithm Architecture
unit, MDU is the motor driver unit, AECU is the acquisition elaboration and
control unit, and PC is the personal computer. The thin straight crack can be described by five parameters
(i.e., x0 , y0 , z0 , a, and l). A previous paper [15] demonstrated
Essentially, there are two types of training, i.e., supervised the close relationship between coordinates x0 and y0 and the
and unsupervised. With the former, data patterns of known field map characteristics (detection and location problems).
identity are repeatedly presented. Provided that the parameters Coordinates x0 and y0 were, therefore, not estimated using the
measured are sufficiently discriminatory, after many iterations, processing algorithm.
the machine will be able to successfully identify classes upon A suitable procedure was set up to perform the evaluation of
which it has been trained. With the latter, on the other hand, a (height), l (length), and z0 (depth), starting from the measured
patterns are presented to the network, and it forms its own field maps. In particular, given that both the complexity and
groupings of the data. the processing time of any regression algorithm (SVM or ANN
Machine learning for classification has two functional steps, in our case) increase with the number of input data, a suitable
i.e., learning and operation. During the learning step, in the feature vector has to be extracted from all the magnetic field
supervised approach, the machine learning system uses a data maps. A preliminary sensitivity analysis was, therefore, carried
set, where each sample is labeled with the appropriate class (the out to select the data-extraction strategy.
so-called ground truth). After the learning step, the system is The first step for the sensitivity analysis is to create a suitable
ready to work with new data (that has never been seen) to solve data set. A completely experimental data set, consisting of EC
the classification problem (operative step). probe data from specimens with known cracks of different
The result of the learning step is a function f (i.e., the shapes and dimensions, would have been too expensive and
classifier) that for a sample x produces as output the belonging time consuming and would have required there to be specimens
class f (x) = C. It also does this on a new sample that is not with known and accurate subsuperficial cracks. An alternative
included in the data set. approach is based on the building of a large data set in a simu-
Machine learning for regression is a system that is capable of lation environment (simulation step), followed by a comparison
identifying a relationship between an independent variable and with experimental data. This validation approach allows a num-
a dependent one. In this case, given a data set of M samples ber of defect configurations to be explored in an easier way.
of x, and the image y produced by the unknown function, The chosen simulation tool is based on an original integral
the result of the learning phase is a function f  that is an formulation of the EC problem [22]. It is able to furnish the
approximation of the function f such that f  (x) ≈ f (x). output of an EC probe when it scans a specimen with a thin
In the whole class of possible defects, researchers have crack, for any crack shape, size, and location, and any probe
concentrated their efforts on the reconstruction of a particu- and specimen characteristics. A 200 mm × 200 mm × 2 mm
lar defect defined as the thin crack. Because of the reduced aluminum plate specimen with a defect region of 20 mm ×
dimensions of typical thin cracks in a conductive material, in 0.01 mm × 2 mm (length × width × height) was considered.
the following, for the sake of simplicity, it is assumed to be a The defect region was divided into a regular grid of 20 × 4
straight crack. In particular, referring to Fig. 2, a thin straight elements (see Fig. 5), imposing defect depth and height steps of
crack is characterized by the following features. 1) It is a 0.5 mm and length step of 1 mm. This way, a suitable data set
bidimensional defect where one dimension (i.e., on the y-axis) of 200 simulated magnetic field maps was obtained.
Authorized licensed use limited to: Centrale Supelec. Downloaded on October 08,2024 at 10:37:32 UTC from IEEE Xplore. Restrictions apply.
1962 IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, VOL. 57, NO. 9, SEPTEMBER 2008

Fig. 5. Defect region considered. An example of a possible crack is highlighted along with its geometrical characteristics.

Fig. 6. Simulated magnetic field amplitude maps of (a) the x-axis sensor and (b) the y-axis sensor. The x- and y-axes in the figures are the scan directions.

Fig. 7. (a) Amplitude of the magnetic field peaks and (b) distance between these peaks versus crack length for different values of crack height and depth
(∗ marker: a = 0.5 mm, z0 = 0 mm; o marker: a = 0.5 mm, z0 = 0.5 mm;
marker: a = 1 mm, z = 0.5 mm; ♦ marker: a = 1 mm, z0 = 1 mm).

To confirm the suitability of this approach, a number of simu- field amplitude maps around the defect area. Similar consider-
lated magnetic field maps were compared to experimental ones. ations are obtained for magnetic phase maps that show huge
For the sake of brevity, Fig. 6 shows the simulated magnetic variations around the defect.
field maps obtained for the same specimen and in the same The following results report the sensitivity analysis for
measurement conditions as the experimental ones reported in parameters a, l, and z0 estimated using simulated magnetic field
Fig. 3. x-axis maps. In this sensitivity analysis, two important variables
Looking at these two figures, it is possible to note that the are considered, i.e., the values assumed by the magnetic field
presence of a defect induces local maxima (peaks) in magnetic amplitude of the peaks and the “peak distance.” The peak
Authorized licensed use limited to: Centrale Supelec. Downloaded on October 08,2024 at 10:37:32 UTC from IEEE Xplore. Restrictions apply.
BERNIERI et al.: CRACK SHAPE RECONSTRUCTION IN ECT USING MACHINE LEARNING SYSTEMS 1963

Fig. 8. (a) Amplitude of the magnetic field peaks and (b) distance between these peaks versus crack height for different values of crack length and depth
(♦ marker: l = 2 mm, z0 = 0 mm; o marker: l = 8 mm, z0 = 0 mm;
marker: l = 20 mm, z0 = 0 mm; ∗ marker: l = 8 mm, z0 = 0.5 mm).

Fig. 9. (a) Amplitude of the magnetic field peaks and (b) distance between these peaks versus crack depth for different values of crack length and height
(♦ marker: l = 2 mm, a = 0.5 mm; o marker: l = 8 mm, a = 0.5 mm;
marker: l = 20 mm, a = 0.5 mm; ∗ marker: l = 8 mm, a = 1 mm).

distance is defined as the distance between the geometrical Similar results are obtained if the maps related to the y-axis
position in the x−y plane of the peaks [for instance, see sensor are considered. For the sake of brevity, these results are
Fig. 3(a)]. Figs. 7–9 show the aforementioned results for crack not reported.
characteristics a, l, and z0 , respectively. Even if the shape of the magnetic field maps and the defect
The following considerations can be made. characteristics (a, l, z0 ) are clearly correlated, the interconnec-
tion among these variables does not allow a simple processing
1) As far as parameter l is concerned, it is strictly related algorithm to be developed.
to the peak distance [Fig. 7(b)], whereas, for l > 5 mm, In light of the results obtained in the preliminary sensitive
the peak amplitude of the magnetic field is not influenced analysis, the intensity and position of the magnetic field peaks
since its values depend only on the pair (a, z0 ) [Fig. 7(a)]. were chosen as the input data for the regression algorithms.
2) The parameter a does not influence the peak distance As described above, the parameter l shows a clear depen-
[see Fig. 8(b)], but it affects the magnetic field amplitude, dence on peak distance; a shows a dependence on l and z0 ; and
which increases for increasing values of a [see Fig. 8(a)]. z0 shows a dependence on l and a. Consequently, the architec-
3) The peak amplitude of the magnetic field decreases with ture of the processing algorithm was designed, considering one
increasing values of z0 [see Fig. 9(a)]. As shown in regressor for each unknown (a, l, z0 ) in a pipelined approach.
Fig. 9(b), for fixed values of the pair (a, l), there is a Fig. 10 shows the whole system, where xil , xia , and xiz0 are
weak relationship between the values of z and the peak the feature vectors that describe, respectively, li , ai , and z0i of
distance. the ith crack in the data set; ˆli , âi , and ẑ0i are the reconstructed
Authorized licensed use limited to: Centrale Supelec. Downloaded on October 08,2024 at 10:37:32 UTC from IEEE Xplore. Restrictions apply.
1964 IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, VOL. 57, NO. 9, SEPTEMBER 2008

3) the maximum acceptable error in parameter reconstruc-

tion as required by typical NDT applications.
As far as point 1) is concerned, the generalization capability
can be defined as the ability to perform well in cases that do not
belong to the training set.
It has been demonstrated [20], [23]–[26] that a good gen-
eralization capability is related to a tradeoff between the
Fig. 10. Architecture of the proposed regressor-based algorithm. maximum reconstruction errors allowed on the training and
test sets.
values; and fˆl , fˆa , and fˆz0 are the reconstructed functions. In particular, the SVR tradeoff is regulated by the
These functions were alternatively reconstructed by means of ε-insensitive parameter (the generalization capability decreases
ANN and SVR to obtain a performance comparison between when this parameter tends to zero, going toward the overfitting
the two machine learning systems. It should be noted that the phenomenon), whereas the generalization capability of the
feature vectors were normalized as required by the processing ANN (where, typically, the higher the number of neurons, the
algorithms. smaller the ANN generalization) is regulated by the number of
SVR and ANN are general-purpose regression techniques epochs during its training phase, its topology, and the number
whose performance, in this case, depends on the following: the of hidden neurons.
kernel selected, ε-insensitive parameter, the kernel parameters Point 2) is investigated by taking into account the effects
for the SVR (see Appendix) and on the network topology, the related to the aspects of the measurement system that are not
activation function, and the number of nodes in the hidden ideal, such as probe noise, the vertical resolution of the data
layer for the ANN. Another important issue for both the ANN acquisition system, the step resolution of the scanning system,
and SVR is the adequacy of the data set (in particular, of the thermal noise, etc.
training set). As for point 3), it is possible to highlight that, in typical NDT
Regarding the SVR, the Radial Basis Function (RBF) kernel applications, adequate defect characterizations are obtained
was adopted since it is particularly suitable to solve these when the accuracy is better than 10%.
kinds of problem [23]. Regarding the ANN, the MultiLayer Taking the above considerations into account, a number of
Perceptron model with the standard Back Propagation learning tests were carried out by implementing the following:
algorithm was adopted. In particular, three-layer feed-forward
1) different ε-insensitive parameter values for the SVR
networks with weights adjusted by conjugate gradient mini-
(varying from 0.0005 to 0.05);
mization were used [24].
2) different numbers of neurons in the hidden layer (n) for
In addition, a tuning phase was required to optimally set
the ANN (varying from 5 to 20);
the other parameters so that both the ANN and the SVR
3) different values of rms noise (k1 ) added onto the esti-
regression algorithms are optimized for the specific application
mated peak distance, where k1 ∈ {0%, 1%, 2%, 5%} is
(i.e., evaluation of a, l, and z0 ).
the rms noise as percentage of the total scan length;
4) different values of rms noise (k2 ) added onto the esti-
III. T UNING OF THE M EASUREMENT M ETHOD mated magnetic field amplitude, where k2 ∈ {0%, 1%,
To deal with the complex dependencies of ANN and SVR 2%, 5%, 10%} is the rms noise as a percentage of the
performance on their parameters, attention was focused on the amplitude of the largest peak.
tuning phase to obtain better accuracy and repeatability in crack Hereinafter, the following notation for rms noises will be
reconstruction. adopted: (k1 , k2 )%, where k1 and k2 have been defined above.
Once the Absolute prediction Error (AE) is defined as AE = For each test, optimum values of our machine learning para-
|Iˆm − Im |, where Iˆm ∈ {ˆl, â, ẑ0 } and Im ∈ {l, a, z0 } are the meters were found, dividing the 200 simulated field maps into
normalized reconstructed parameters and the normalized refer- 170 maps per training set and 30 maps per test set.
ence, respectively, the following figures of merit were consid- For the sake of brevity, only a few tests are reported. Fig. 11
ered and minimized: shows a comparison of the reconstruction performances of the
1) the Mean Absolute prediction Error (MAE), defined as SVR and ANN processing algorithms. It shows, first, the MAE

behavior for the three SVRs (a, l, and z0 ) versus both noise level
MAE = 1/N N ˆ
m=1 |Im − Im |, where N is the cardinal-
ity of the test set; and three different ε-insensitive values (ε = 0.0005, ε = 0.005,
the standard deviation of AE, i.e., σAE =
2)  and ε = 0.05) and, second, the MAE behavior for the three

ANNs (a, l, and z0 ) versus both noise level and four different
1/N N (Iˆm − Im )2 .
m=1 numbers of neurons in the hidden layer (n = 5, n = 10, n =
The tuning phase was carried out by taking into account the 15, n = 20). The following observations can be made.
following three important aspects: 1) For all noise levels, ε-insensitive values, and number
1) the SVR and ANN generalization capability; of nodes imposed, the SVR generally exhibits better
2) the SVR and ANN robustness in relation to superimposed MAE values than the ANN. In particular, for the SVR
noise (to take into account the noise present on the (see Table I), the best case was found in the absence
experimental data); of noise ((k1 , k2 ) = (0, 0)%), whereas the worst case
Authorized licensed use limited to: Centrale Supelec. Downloaded on October 08,2024 at 10:37:32 UTC from IEEE Xplore. Restrictions apply.
BERNIERI et al.: CRACK SHAPE RECONSTRUCTION IN ECT USING MACHINE LEARNING SYSTEMS 1965

Fig. 11. MAE values versus rms noise for (a) length, (b) height, and (c) depth. The crack characteristics are alternatively reconstructed using SVR and ANN.

was obtained with the maximum considered level of TABLE I

BEST AND WORST CASE FOR THE SVR IN TERMS OF MAE
noise ((k1 , k2 ) = (5, 10)%), both with the same value for AND σAE FOR l, a, AND z0 W ITH R ESPECT TO THE
ε = 0.0005. We can observe that the same value for ε ADDED NOISE AND THE ε PARAMETER
produces the best and the worst case, depending on the
superimposed noise.
2) It is possible to observe that when the noise levels are
low (nearing ideal measurement conditions), the MAE
and ε are directly proportional. This means that high
specialization (low value of ε) produces a reduction in
MAE. On the contrary, when the noise levels are high
(nearing real measurement conditions), the MAE and 5) Table I also shows the values of σAE for both the best
ε are inversely proportional. A low level of specializa- and the worst cases using the SVR. Similar results are
tion (high value of ε), therefore, produces an increase obtained for the other cases. The analysis of this figure of
in MAE, which means that a high level of general- merit confirms the stability of the developed reconstruc-
ization produces a prediction system that is substan- tion machines.
tially immune to noise and, therefore, suitable for a real These observations show that the SVR trained with ε = 0.05
situation. and (k1 , k2 ) = (1, 1)%, and the ANN trained with n = 5 and
3) For the ANN, there is no evident relation between the (k1 , k2 ) = (1, 1)%, are probably the best tradeoff between the
MAE and the number of hidden nodes. required measurement accuracy and the generalization capabil-
4) Considering 1% as a typical accuracy of a good industrial ity. Thus, this pair of machines was chosen to be adopted for
measurement station, we will adopt the pairs (k1 , k2 ) = the defect reconstruction of the experimental data. The results
(1, 1)% for noise level in the training phase. will be shown in the next section.
Authorized licensed use limited to: Centrale Supelec. Downloaded on October 08,2024 at 10:37:32 UTC from IEEE Xplore. Restrictions apply.
1966 IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, VOL. 57, NO. 9, SEPTEMBER 2008

TABLE II In addition, in the experimental tests, the ANN method gave

GEOMETRICAL CHARACTERISTICS OF THE DEFECTS USED ON AN
ALUMINUM PLATE OF 200 mm × 200 mm × 2 mm worse results compared to the SVR one. In particular, large
values of MAE were obtained on the experimental test set
considered (MAEl = 0.14, MAEa = 0.24, MAEz0 = 0.24).
As for the SVR, an acceptable reconstruction performance
was obtained. In particular, good results are obtained on the
parameter l with a maximum error equal to 0.8 mm (on a crack
with l = 5 mm) and MAEl = 0.043.
A lower relative reconstruction quality is obtained on
height (a) and depth (z0 ) parameters with maximum errors
equal to 0.5 mm (on a crack with a = 2 mm) and 0.3 mm
(on a crack with z0 = 0 mm), respectively, and with MAE
values MAEa = 0.12 and MAEz0 = 0.12.
The most critical results are obtained in the reconstruction
of the height parameter. Generally, the overall reconstruction
performance deteriorates, reducing the dimensions of the crack
(i.e., length and height).
Finally, these results also confirm the adequacy of the data set
and the suitability of the chosen tradeoff between specialization
and generalization.

V. F INAL R EMARKS
This paper has described a measurement method for defect
characterization in NDT of conductive materials based on the
ECT approach. It is based on an NDT measurement station
and a suitable processing procedure using machine learning
systems.
The comparison between two machine learning systems [an
ANN and a SVM for Regression (SVR)], both in a simulated
environment and in real cases, proved that the SVR is better for
solving crack characterization problems in ECT.
The experimentation carried out to validate the results
achieved highlights the suitability of the SVR-based approach,
confirming the accuracy and repeatability of the results ob-
tained in defect characterization for known real cases. The
reconstruction error, which is similar for all crack dimensions,
is already acceptable for length, although it still needs to be
Fig. 12. Comparison between reference and reconstructed l, a, and z0 pa- improved for height and depth.
rameters for rms noise (k1 , k2 ) = (1, 1)% alternatively using ANN (with It is supposed that a further improvement in the results can
n = 5) and SVR (with ε = 0.0005).
be obtained through a training phase carried out by suitably
enlarging the training set by both increasing the resolution
IV. E XPERIMENTAL V ALIDATION
step in crack dimension definition and making the kind of
A number of experimental tests were carried out on the mea- defect more general by removing the straight crack hypothesis.
sured field maps obtained by scanning 200 mm × 200 mm × In fact, it is necessary to remark that, although in some real
2 mm aluminum specimens with known defects (Table II). applications, the cracks are well defined and very similar to the
Each experimental field map was obtained using a biaxial straight crack considered, this hypothesis does not fit with all
fluxset probe (20 pickup turns and 30 driving turns) and an classes of real defects. These can be nonlinear and nonuniform
excitation coil (100 turns, 4 mm cylindrical internal radius either in thickness or in depth or can have multiple branches. In
support, 22 mm height) supplied with 500 mA, 5 kHz sinu- these cases, despite the good generalization capability shown
soidal current. Fluxset driving coils were supplied with 44 mA, by the developed SVR measurement software, the proposed
50 kHz triangular current, and the pick-up coil signals were system should be unable to overcome these difficulties. Re-
sampled at 10 MHz, collecting two 100k point buffers. garding the nonuniformity both in thickness and depth, this can
The results obtained using the above-described tradeoff (the be solved by enlarging the numbers of parameters that need
ANN trained with (k1 , k2 ) = (1, 1)% rms noise and n equal to be reconstructed (i.e., both initial and final thickness and
to 5 and the SVR trained with (k1 , k2 ) = (1, 1)% rms noise and depth). Obviously, it is, consequently, necessary to enlarge the
0.05 ε-insensitive value) are shown in Fig. 12 and described in training set by also considering magnetic field maps related
the following. to this kind of defect. As for both the nonlinearity and the
Authorized licensed use limited to: Centrale Supelec. Downloaded on October 08,2024 at 10:37:32 UTC from IEEE Xplore. Restrictions apply.
BERNIERI et al.: CRACK SHAPE RECONSTRUCTION IN ECT USING MACHINE LEARNING SYSTEMS 1967

presence of multiple branches, it is a more difficult problem estimate a function f that minimizes R[f ]. Formally, it leads to
to overcome because it is not easy to define the parameters a convex optimization problem
to be reconstructed. The solution to these problems requires 
further research activity that is already in process. It aims to 1 yi w, φ(xi ) − b ≤ ε
Maximize w 2 subject to (4)
extend the proposed approach to volumetric defect. Branched 2 w, φ(xi ) + b − yi ≤ ε.
and nonlinear defects can thus be addressed as a suitable region
Sometimes, such a function does not exist, or it is possible to
with an equivalent regular shape.
tolerate some errors. In the latter case, it is possible to introduce
slack variables ξ, ξ ∗ , and the problem formulation becomes [23]
A PPENDIX 
t
SVM Maximize 1/2 w 2
+C (ξ + ξ ∗ )
i=1
SVM was introduced in 1992 [25] as a pattern classification

and regression technique. As for the SVR [23], this structure,  yi − (w, φ(xi ) + b) ≤ ε + ξ ∗
like the ANN, is a very powerful learning machine and can subject to −yi + (w, φ(xi ) + b) ≤ ε + ξ (5)
represent arbitrary functions. To highlight the differences be- 
ξ, ξ ∗ ≥ 0
tween an SVR and an ANN, the ANN training phase employs
an empirical risk minimization principle, which minimizes the where i = [1, . . . , t], and C is the tradeoff between the flatness
error on the training data, whereas the SVR adopts a structural of f and the amount of points that are tolerated out of the
risk minimization principle, which minimizes the upper bound insensitive tube.
on the generalization error. There are, therefore, a number of The optimization problem above can more easily be solved
difficulties inherent to ANN design, namely model selection in its dual formulation. It is possible to construct a Lagrange
and parameter settings (generally, these choices are a result of function from both the objective function (the so-called primal
designer experience and empirical considerations). The SVR, objective function) and the corresponding constraints by intro-
instead, is trained by solving a constrained quadratic optimiza- ducing a dual set of variables. It can be shown that this function
tion problem, and to identify the optimal architecture as well as has a saddle point in relation to primal and dual variables at the
evaluate the performance influence of some design parameters, optimal solution. After several reductions, we obtain the dual
several SVR setups have to be tested. optimization problem
Let the training data be X = {(x1 ,y1 ),. . . ,(xt , yt )} ⊂ χ×R,

where xi ∈ χ ⊂ Rd is the input pattern, yi ∈ R is the target 

t

 −2
1
(αi − αi∗ ) αj − αj∗ k(xi , xj )
value, and t is the total number of training samples. Further- i,j=1
more, assume that this training set has been drawn indepen- Maximize



t

t
dently and identically distributed with a probability distribution  −ε (αi + αi∗ ) + yi (αi + αi∗ )
i=1 i=1
P (x, y). In ε-SVR, the goal is to find a function f (x) that
has, at most, ε deviation from the actually obtained targets 

t
yi on the training set. The general form of the prediction (αi − αi∗ ) = 0
subject to (6)
function is  i=1
αi , αi∗ ∈ [0, C].
f (x) = w, φ(x) + b (1)
Solving (6) with related constraints,

we determine the Lagrange
where ·, · is the dot product in Ω, i.e., a feature space with multiples α and α∗ and obtain w = ti=1 (αi + αi∗ )φ(xi ).
a dimension that is generally different from χ such that φ : Therefore, the estimation function in (1) becomes
χ → Ω, and b ∈ R. The goal is to determine w and b, starting 
t
from training data through the minimization of a functional f (x) = (αi − αi∗ ) φ(x), φ(xi ) + b. (7)
risk [26] i=1
 The computation of b can be done by exploiting the so-called
R[f ] = c (x, y, f (x)) dp(x, y) (2) Karush–Kuhn–Tucker conditions, according to which, the
product of dual variables and constraints has to vanish at
based on the empirical data X, where c(x, y, f (x)) is a cost the optimal solution point. The examples (points belonging to
function that determines how we will penalize estimation er- the training test) that come from nonvanishing coefficients are
rors. In general, the so-called ε-insensitive is used as a loss called Support Vectors.
function By using the kernel trick, (7) can be written as
 
t
0, if |y − f (x)| < ε
c (x, y, f (x)) = (3) f (x) = (αi − αi∗ ) k(x, xi ) + b (8)
|y − f (x)| − ε, otherwise
i=1

that defines the insensitive tube around f (x). Since the proba- where the kernel function k(x, xi ) = φ(x), φ(xi ) is a sym-
bility measurement dp(x, y) is unknown, we can only use X to metric function. Mercer’s theorem [23] ensures the availability
Authorized licensed use limited to: Centrale Supelec. Downloaded on October 08,2024 at 10:37:32 UTC from IEEE Xplore. Restrictions apply.
1968 IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, VOL. 57, NO. 9, SEPTEMBER 2008

of such functions, which are called kernels k(xi , xj ). [23] J. A. Smola and B. Scholkopf, “A tutorial on support vector regression,”
Some of the kernels frequently used in SVM applications Roy. Holloway College, London, U.K., NueroCOLT Tech. Rep. NC-TR-
98-030, 1998.
(both for classification and for regression) are polynomial [24] R. Rojas, Neural Networks, A Systematic Introduction. New York:
k(xi , xj ) = (xi , xj  + c)d , Gaussian (RBFs) k(xi , xj ) = Springer-Verlag, 1996, ch. 7, pp. 151–184.
exp(−(1/δ 2 )(xi − xj )2 ), and sigmoid k(xi , xj ) = [25] B. E. Boser, I. M. Guyon, and V. N. Vapnik, “A training algorithm for op-
timal margin classifiers,” in Proc. 5th Annu. Workshop Comput. Learning
tanh(βxi , xj  + γ), where c, d, δ, β, and γ are kernel Theory, Pittsburgh, PA, 1992, pp. 144–152.
parameters. [26] V. N. Vapnik and A. Lerner, Estimation of Dependences Based on Empir-
ical Data. Berlin, Germany: Springer-Verlag, 1982.

R EFERENCES
Andrea Bernieri was born in Napoli, Italy, in 1956.
[1] EU Directive on the responsibility for defective products, 1999. Directive He received the M.S. and Ph.D. degrees from the
85/374/EEC, amended by Directive 99/34/EC. University of Napoli in 1982 and 1987, respectively,
[2] “NDT handbook,” in Electromagnetic Testing, 2nd ed., 4b. P. McIntyre both in electrical engineering.
and M. Mester, Eds. Columbus OH: Amer. Soc. Nondestructive Testing, Since 1989, he has been with the Department
1986. of Industrial Engineering, University of Cassino,
[3] D. C. Jiles, “Review of magnetic methods for non-destructive evaluation,” Cassino, Italy, where he was an Assistant Professor
NDT Int., vol. 21, no. 5, pp. 311–319, 1988. of electrical measurements in 1989, was an Associate
[4] D. C. Jiles, “Review of magnetic methods for non-destructive evaluation Professor of electrical measurements in 1998, and
(Part 2),” NDT Int., vol. 23, no. 2, pp. 83–92, Apr. 1990. has been a Full Professor of electrical measure-
[5] R. K. Chapman et al., “IEE Colloquium on new applications in modelling ments since 2002. He is principally concerned with
and inversion techniques for non-destructive testing,” in Proc. IEE research into artificial intelligence techniques for model identification, fault
Colloq. New Appl. Modelling Inversion Tech. Non-Destructive Testing, detection, and diagnosis applied to electrical component and systems.
1999, pp. 1–38.
[6] S. Calcagno et al., “Damage analysis of thin metallic plates,” in Proc.
ASEM, Pusan, Korea, 2002.
[7] Z. Chen, K. Aoto, and K. Miya, “Reconstruction of cracks with physical Luigi Ferrigno (M’04) was born in Italy in 1972. He
closure from signals of eddy current testing,” IEEE Trans. Magn., vol. 36, received the M.S. degree in electronic engineering
pt. 1, no. 4, pp. 1018–1022, Jul. 2000. from the University of Salerno, Salerno, Italy, in
[8] G. Preda, B. Cranganu-Cretu, F. I. Hantila, O. Mihalache, Z. Chen, and 1998 and the Ph.D. degree in electrical engineering
K. Miya, “Nonlinear FEM-BEM formulation and model-free inversion from the University of Napoli, Napoli, Italy, in 2002.
procedure for reconstruction of cracks using pulse eddy currents,” IEEE Since 2001, he has been an Assistant Professor
Trans. Magn., vol. 38, no. 2, pp. 1241–1244, Mar. 2002. of electrical and electronic measurements with the
[9] S.Calcagno, F. C. Morabito, and M. Versaci, “A novel approach for de- University of Cassino, Cassino, Italy. Since 2004,
tecting and classifying defects in metallic plates,” IEEE Trans. Magn., he has also been the Head of the Laboratory of
vol. 39, pt. 1, no. 3, pp. 1531–1534, May, 2003. Industrial Measurement, University of Cassino. His
[10] Y. Nagaya, T. Takagi, T. Uchimoto, and H. Huang, “Identification of mul- current research interests are concerned with the
tiple cracks from eddy-current testing signals with noise sources by image measurement system for nondestructive testing via eddy currents, wireless
processing and inverse analysis,” IEEE Trans. Magn., vol. 40, pt. 2, no. 2, systems and wireless sensor network realization, and characterization, methods,
pp. 1112–1115, Mar. 2004. and instruments for the characterization of passive components and ac mains
[11] H. Huang, T. Takagi, and T. Uchimoto, “Crack shape reconstruction under nonsinusoidal conditions.
in ferromagnetic materials using a novel fast numerical simulation
method,” IEEE Trans. Magn., vol. 40, pt. 2, no. 2, pp. 1374–1377,
Mar. 2004. Marco Laracca (M’06) was born in Formia, Italy, in
[12] Y. Nagaya, H. Endo, T. Uchimoto, T. Takagi, H. J. Jung, and 1973. He received the M.S. degree in electrical engi-
L. Bartolomeo, “Crack shape reconstruction using ECT camera system,” neering and the Ph.D. degree in electrical and infor-
in Proc. SICE Annu. Conf., Sapporo, Japan, 2004, vol. 3, pp. 2760–2763. mation engineering from the University of Cassino,
[13] J. Cox and D. Brown, Eddy Current Instrumentation. Snoqualmie, WA: Cassino, Italy, in 2002 and 2006, respectively.
Zetec Inc., 2003. Since February 2006, he has been an Assis-
[14] A. Bernieri, L. Ferrigno, M. Laracca, and M. Molinara, “An SVM ap- tant Professor of electrical and electronic measure-
proach to crack shape reconstruction in eddy current testing,” in Proc. ments with the University of Cassino. His current
IMTC, Sorrento, Italy, 2006, pp. 2121–2126. research interests are concerned with the realization
[15] A. Bernieri, G. Betta, L. Ferrigno, and M. Laracca, “A bi-axial probe for of measurement systems for nondestructive testing
non destructive testing on conductive materials,” IEEE Trans. Instrum. via eddy currents, sensor realization and character-
Meas., vol. 53, no. 3, pp. 678–684, Jun. 2004. ization, power quality measurements, and electric power measurement under
[16] G. Betta, L. Ferrigno, and M. Laracca, “Calibration and adjustment of nonsinusoidal conditions.
an eddy current based multi-sensor probe for non-destructive testing,” in
Proc. SICON, Houston, TX, 2002, pp. 120–124.
[17] A. Bernieri, G. Betta, and L. Ferrigno, “Characterization of an eddy-
current-based system for non-destructive testing,” IEEE Trans. Instrum. Mario Molinara received the “Laurea” degree with
Meas., vol. 51, no. 2, pp. 241–245, Apr. 2002. honors in computer science from the University
[18] A. Bernieri, G. Betta, G. Rubinacci, and F. Villone, “A measurement of Sannio, Benevento, Italy, in 1999 and the Ph.D.
system based on magnetic sensors for non destructive testing,” IEEE degree in computer science and telecommunication
Trans. Instrum. Meas., vol. 49, no. 2, pp. 455–459, Apr. 2000. from the University of Salerno, Salerno, Italy,
[19] A. Bernieri, G. Betta, L. Ferrigno, and M. Laracca, “An automated in 2003.
self-calibrated instrument for nondestructive testing on conductive ma- In 2004, he joined the Dipartimento di Automa-
terials,” IEEE Trans. Instrum. Meas., vol. 53, no. 4, pp. 955–962, zione, Elettromagnetismo, Ingengeria dell’Infor-
Aug. 2004. mazione e Matematica Industriale, University of
[20] R. O. Duda, P. E. Hart, and D. G. Stork, Pattern Classification, 2nd ed. Cassino, Cassino, Italy, where he is currently a Re-
Hoboken, NJ: Wiley, 2001, pp. 282–350. searcher in computer science. His current research
[21] D. E. Rumelhart and J. L. McClelland, PDP Research Group, Parallel interests include image analysis and interpretation, classification techniques,
Distributed Processing, vol. 1, 1996, Cambridge, MA: MIT Press. statistical learning, biomedical imaging, neural networks, optical character
[22] R. Albanese, G. Rubinacci, and F. Villone, “An integral computational recognition, and map and document processing.
model for crack simulation and detection via eddy currents,” J. Comput. Dr. Molinara is a member of the International Association for Pattern
Phys., vol. 152, no. 2, pp. 736–755, Jul. 1999. Recognition (IAPR).
Authorized licensed use limited to: Centrale Supelec. Downloaded on October 08,2024 at 10:37:32 UTC from IEEE Xplore. Restrictions apply.