Biomaterial Crack Geometry Estimation Using Eddy

Current Inverse Approach

Milan Smetana, Lukas Behun, Daniela Gombarska, Ladislav Janousek

Department of Electromagnetic and Biomedical Engineering
University of Zilina, Faculty of Electrical Engineering
Zilina, Slovak Republic
milan.smetana@fel.uniza.sk

Abstract—The main subject of this paper is inverse problem knowledge of the gradient of the error with respect to a
approach to non-destructive evaluation of conductive austenitic variation of the crack geometry.
stainless steel as biomaterials. Eddy current evaluation under This article deals with geometry estimation of the defined
harmonic excitation is used. Magneto-resistive effect sensors are artificial cracks given from ECT response signals. The
there in the role of pick-up elements. Basic cracks geometry is algorithm for an inverse problem solution is based on the
inversely estimated based on computational algorithms. The
Wavelet transform and on the artificial neural network.
gained results are presented and discussed in the paper.
Obtained results are presented and discussed.
Keywords—Inverse problem, non-destructive evaluation, wavelet
transform, biomaterial cracks, principal component analysis,
II. EDDY CURRENT TESTING METHOD
Bayesian regularization back-propagation. One of the conventional electromagnetic methods,
originating from the electromagnetic induction phenomena, is
I. INTRODUCTION eddy current non-destructive inspection. This method is
A role of non-destructive evaluation of material structures theoretically well known and widely utilized in the practice. It
is of global importance. Periodic inspection of components is suitable for evaluating surface, subsurface and near surface
cracks and it is applicable for almost all materials with non-
and devices ensures their safe, effective and long-term
zero conductivity. Numerous advantages such as high
operation. New methods and devices are still being developed
sensitivity, rapid scanning, contactless inspection, and
and designed to tackle gradually increasing demands for versatility contribute to its wide utilization. The evolution of
reliable detection and precise characterization of material the computer systems and numerical methods, expanded the
discontinuities. interest in ECT as problems that had been difficult to solve can
Eddy current non-destructive techniques are widely used in now be dealt by numerical means. New technologies empower
the field of electromagnetic inspection of metallic structures to to very precisely and rapidly perform forward numerical
determine their composition and to detect possible cracks. simulations concerning different applications of the ECT. The
These techniques find widespread application in industry. The major characteristics of conventional ECT with harmonic
growth has been fueled by an increasing ability to extract driving, [5], [6], [7] are: application of discrete frequencies,
information from the signal generated by the probe, [1], [2]. large dynamic range, accurate crack sizing, high speed, signal
Eddy‐current inversion is the process whereby the geometry of analysis by pattern recognition.
a crack in a metal is derived from electromagnetic probe The basic principle of the ECT is quite simple: it is based on
measurements. The problem can be described as the task of the electromagnetic induction phenomenon, Fig. 1. A coil
reconstructing an unknown distribution of electrical driven with a time-varying current generates the time varying
conductivity from eddy‐current probe impedance electromagnetic field (EMF) in its vicinity. Due to EMF, time-
measurements recorded as a function of probe position, varying electromotive force is induced in adjacent conductive
excitation frequency or both of them. This is widely materials. As a consequence, eddy currents flow in the object
recognized as a central theoretical problem whose solution is according to the electromotive force. The EMF generated by
likely to have a significant impact on the characterization of the eddy currents has the opposite direction in comparison
cracks in conducting materials. An inversion scheme is thus with the exciting EMF generated by the coil. In case of a crack
described for finding the shape and size of cracks from eddy‐ presence, it influences the flow pattern of the induced eddy
current probe impedance measurements. The approach is currents. The impedance of the coil changes due to this fact,
based on an optimization scheme that seeks to minimize a giving the means for obtaining information about the material
global error function quantifying the difference between flaw, [1], [2], [3], [4], [8].
predicted and observed probe impedances. The error minimum
is sought using a standard descent algorithm, requiring
 Where A [T.m] is the magnetic vector potential and V [V]
is the electric scalar potential. The ECT analysis is conducted
by both of these quantities.
The solution domain is subdivided into conducting area Ω1
and non-conducting area Ω2. The eddy currents in conductor
are governed by the following equations in:

air region: 2 A = 0 (11)

coil region:  2 A = -μJ (12)

conductor region:  σ(-  V - jωA) = 0 (13)

Fig. 1. Principle of eddy current method, [1].  2 A - jωA- μσ  V = 0 (14)

The ECT can be modeled using the quasi-stationary EMF The EMF analysis is based on the magnetic vector potential
approach. Usually, this approach gives reliable results when the formulation (MVP). The set of the equations (11) to (14) is
time changes of EM field are relatively slow, so the usually solved by numerical means using the finite element
displacement current can be neglected, applies . This method (FEM).
EMF condition applies in conductive materials even at higher III. INVERSE ECT APPROACH
frequencies because the conductive current is much higher than
the displacement current there. The quasi-stationary EMF is Solution of an inverse problem holds still the current issue
described by the four Maxwell’s equations in the following status. Huge computational power of modern computers
form: allows solving of very complex and nonlinear tasks. Actual
main aim is to obtain relevant crack geometry and/or to
curl H = J , (1) visualize 3D crack profile in real time. This study shows the
results based on algorithm developed on the basis of the
curl E = - , (2) wavelet transform (WT) and the artificial neural network. The
WT is used for determination of width and length of the
div B = 0 , (3) inspected cracks. Because of its properties (non-symmetrical
div D = ρ0 , (4) wave that is suitable for detection of a gradient) the Haar
mother wave is used. Computation of the gradient is based on
where: H [A.m-1] is the magnetic intensity vector, a provisional crack estimate and then the crack estimate is
E [V.m-1] is the electric intensity vector, B [T] is the magnetic updated in a direction that reduces the error. The process
flux density vector, D [A.s.m-2] is the electric displacement continues iteratively until a convergence criterion has been
vector, J [A.m-2] is the conducting current density vector and satisfied, [3], [4]. The convolution process of the input data
ρ0 [C.m-3] is the volume density of a free charge. The material matrix is used. Computed and weighted WT coefficients are
relations valid for vector quantities of EM field are in case of directly used to determine the crack geometry as well as its
linear, homogeneous and isotropic environment in the spatial orientation. The crack depth estimation, based on the
following form: artificial neural network, uses so-called data reduction
approach: the principal component analysis (PCA) method
D = ε E, (5) was applied and the Bayesian regularization back-propagation
(BRB) is used as the training function. BRB is a network
B = μ H, (6) training function that updates the weight and the bias values
J = σ E, (7) according to Levenberg-Marquardt optimization. It minimizes
a combination of squared errors and weights, and then
Where ε [F.m-1] is the permittivity, μ [H.m-1] is the determines the correct combination so as to produce a network
permeability and σ [S.m-1] is the conductivity of a material. that generalizes well. The process is called Bayesian
The EM field can be analyzed using the potential functions: regularization [5], [9], [10], [11], [12]. These measured data
are reduced from the matrix form into the vector one by use of
B = curl A, (8) PCA method and these vectors are also used for the training
process.
grad V= - E – , (9)

div A = 0 , (10)
 IV. EXPERIMENTAL SET-UP V. RESULTS AND DISCUSSION
For the purposes of this study, a sample of conductive The computed results of the inverse ECT solution are
plate with thickness h = 10 mm and electromagnetic presented in this section. All the gained signals composed an
parameters of AISI 316L stainless steel is utilized. Considered input 3D matrix to be used for computational purposes.
AISI 316L stainless steel material properties are conductivity Application of stated mathematical algorithms was then
of σ = 1.4 MS/m and the relative permeability of μr = 1. The performed. The output of these procedures is in the form of
specimen contains five non-conductive electric-discharge- discrete numbers that represent estimated crack geometry.
machined (EDM) cracks with the rectangular shape. Basic parameters are thus inversely estimated.
Dimensions of plate cracks declared by the manufacturer are Columns GMR and AMR in Table I present computed
summarized in Tab. I. results that reflect the estimation of the basic geometry of the
The used ECT probe consists of two exciting coils that are inspected cracks in comparison with their real dimensions
positioned normally to the surface of the inspected material listed in column OD. Computed values of crack dimensions
apart from each other, Fig. 2. only slightly differ from OD in both utilized magnetometers.
This indicates suitability of these sensors for such type of
TABLE I. INSPECTED EDM CRACKS GEOMETRY: REAL AND COMPUTED cracks under harmonic excitation with given, relatively low
VALUES excitation frequency.
Crack Length [mm] Width [mm] Depth [mm] Table II display the relative errors computed for both
No. OD *
GMR AMR OD *
GMR AMR OD *
GMR AMR magnetometers individually. It can be observed that the most
1 3 4 3.8 0.2 0.38 0.47 1 1.1 0.95 important information about the cracks geometry (depth
2 9 10.1 10.2 0.2 0.39 0.59 3 3 3.2 parameter) is in good correlation with the real values. On the
3 15 16 16.1 0.25 0.4 0.6 5 5.6 5.5 other hand, results obtained for width estimation are of high
4 21 22.5 22.5 0.25 0.4 0.6 7 6.3 6.1 dispersion. This is caused by the dimensions (effective sensing
5 27 25.5 26 0.25 0.6 0.6 9 8.1 8.5
*
area and thus resolution) of the pick-up elements and also by
OD stands for original dimension of flaw given in plate datasheet
the excitation frequency. It could be said that crack width can
be evaluated by additional method (e.g. visual inspection). On
The coils are connected in series but magnetically opposite the field of ECT practice, the most important crack information
to decrease the coupling between the exciting system and the is about its depth and length. The width parameter is never
sensing element. High sensitivity of a pick-up circuit can be primarily evaluated. Proposed method showed close estimation
adjusted in such case. The exciting coils with self-inductances of these parameters and its precision can be seen.
of L1 = 2.24 mH, L2 = 2.23 mH are driven by the harmonic
current with an effective value of I = 0.85 A and a frequency of TABLE II. OBTAINED RESULTS: INSPECTED CRACKS, WIDTH
f = 1 kHz, Fig. 2. PARAMETER ESTIMATION
Crack L GMR L AMR W GMR W AMR D GMR D AMR
No. [%] [%] [%] [%] [%] [%]
1 3 4 3.8 0.2 0.38 0.47
2 9 10.1 10.2 0.2 0.39 0.59
3 15 16 16.1 0.25 0.4 0.6
4 21 22.5 22.5 0.25 0.4 0.6
5 27 25.5 26 0.25 0.6 0.6

Obtained results in the graphical form display Fig. 3 to

Fig.5. Matching of the computed values with OD is clearly
presented. Given figures imply that the resolution power of
Fig. 2. Configuration of the ECT probe above the inspected material. used magnetometers is sufficient. Used computational means
showed an effective tool for such purpose.
As the pick-up elements are used commercially available
GMR (GF708, Sensitec GmbH) and AMR (AF755B, Sensitec
GmbH.) magnetometers, respectively. The pick-up sensors are
oriented normally to the surface of the inspected specimen. In
such geometry is sensed the perpendicular component of the
magnetic field. Each of the cracks is inspected (near side
approach) two-times with both the sensing elements,
respectively. The signals are measured of scanning area with
the dimensions of s = 60 mm x 60 mm, relative to the crack
center (Fig.4). The lock-in amplifier is used for filtering of the
gained signals. Measured data are acquired using data Fig. 3. Numerical results: crack length estimation
acquisition card with resolution of 16bits/channel, 15kS/sec.
User interface for data manipulation, controlling the stage and
processing the data is created using the LabVIEW development
environment.
 VI. CONCLUSION
This paper proposed solution of inverse problem in
electromagnetic non-destructive evaluation. Two
magnetometer types of sensing probes were used along with
pancake excitation coils under the harmonic excitation for eddy
current evaluation: AMR and GMR magnetometers. The
measured response signals obtained by two-dimensional
scanning of the austenitic steel plate with presence of five
artificial cracks were used as the input data for proposed
Fig. 4. Numerical results: crack width estimation
solution. The developed algorithm for crack length, width and
depth estimation was based on the Wavelet transform along
with the artificial neural network. The software solution was
implemented by means of the MATLAB software. The output
of this solution was data of the cracks geometry which were
compared with the datasheet values. The relative error was
computed for each data set. Obtained results showed that
proposed software solution estimated results are in good
correlation with the datasheet values. Future work of the
authors will be addressed to optimization of the algorithm
focused on more precise estimation of sub-surface as well as of
Fig. 5. Numerical results: crack depth estimation the fatigue defects.
REFERENCES
[1] Behun, L., Smetana, M.:Decreasing Uncertainty in Width Estimation of
EDM Cracking from Eddy-Current Signals, In ELEKTRO 2016: 11th
international conference: proceedings. Štrbské Pleso - High Tatras:
IEEE, 2016. ISBN 978-1-4673-8698-2. - CD-ROM, p. 474-477.
[2] Chen, Z., Yusa, N.: Inversion techniques for eddy current NDE using
optimization strategies and a rapid 3D forward simulator, International
Journal of Applied Electromagnetics and Mechanics 20 (2004) s.179–
187, IOS Press.
Fig. 6. Numerical results: length estimation, relative error dependence. [3] Hwang, B., Hillmann, S., Schulze , M., Klein, M, Heuer, H.: Eddy
current imaging for electrical characterization of silicon solar cells and
TCO layers, Proceedings Volume 9439, Smart Materials and Non-
destructiveEvaluation for Energy Systems, 2015; 94390D (2015).
[4] Janousek, L., Rebican, M., Smetana, M., Duca, A.: Progress in Three-
Dimensional Non-Destructive Assessment of Real Cracks from Eddy
Current Testing Signals, Communications, 1A/2015, Volume 17.
[5] Karban, P., Kotlan, V., Dolezel, I.: Numerical model of induction shrink
fits in monolithic formulation, IEEE Transactions on Magnetics, 2012,
Volume 48, Issue 2, pp. 315-318, ISSN: 00189464
[6] Karban, P., Mach, F., Kus, P., Panek, D., Dolezel, I.: Numerical solution
of coupled problems using code Agros2D, Computing, 2013, Volume
Fig. 7. Numerical results: width estimation, relative error dependence. 95, Issue 1 supplement, pp. 381-408, ISSN: 0010485X.
[7] Pirani, A., Ricci, M., Tamburrino A., Ventre S.: 3D reconstruction of
cracks in metallic materials by eddy current isnpections, In
Electromagnetic Non-destructiveEvaluation XI, IOS Press, 2008, Studies
in Applied Electromagnetics and Mechanics 31, ISSN 1383-7281.
[8] Ptak, P., Borowik, L.: Analysis of impact of frequency on sensitivity of
inductive sensor for measuring thickness of coatings, Electrical Review,
Vol. 89, Issue 12, 2013, Pages 269-271.

Fig. 8. Numerical results: depth estimation, relative error dependence.

Figures 6-8 show values of computed relative errors for

individual defects represented by crack number on horizontal
axis. The best achieved results were obtained for a critical
parameter: crack depth. The relative error for crack width
estimation oscillates while its value for length estimation
exponentially decreases.