See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/262771499

Automatic Dynamic Depth Focusing for NDT

Article in IEEE Transactions on Ultrasonics Ferroelectrics and Frequency Control · April 2014
DOI: 10.1109/TUFFC.2014.2955

CITATIONS READS
35 726

4 authors:

Jorge Camacho Jorge F. Cruza

Tecnológico de Monterrey Spanish National Research Council
75 PUBLICATIONS 1,113 CITATIONS 34 PUBLICATIONS 303 CITATIONS

SEE PROFILE SEE PROFILE

Jose Brizuela Carlos Fritsch

National Scientific and Technical Research Council Spanish National Research Council
28 PUBLICATIONS 263 CITATIONS 89 PUBLICATIONS 1,425 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Jose Brizuela on 15 October 2018.

The user has requested enhancement of the downloaded file.

IEEE Transactions on Ultrasonics, Ferroelectrics, and Frequency Control , vol. 61, no. 4, April 2014 673

Automatic Dynamic Depth Focusing for NDT

Jorge Camacho, Jorge F. Cruza, Jose Brizuela, and Carlos Fritsch, Senior Member, IEEE

Abstract—Auto-focusing along with dynamic depth fo- cal laws can be pre-computed and stored in the equip-
cusing (DDF) would be very valuable to inspect arbitrarily ment, as is usually done in medical echography.
shaped parts when operating with wedges or with other cou-
pling media to avoid the burden of computing and setting the
However, for arbitrarily shaped parts or, more gener-
correct focal laws while still getting the best possible resolu- ally, when a coupling medium is located between the ar-
tion at all depths. ray probe and the inspected part, refraction effects must
This work proposes a three-step procedure to perform the be considered. When the geometry is relatively constant,
auto-focusing function with DDF in real time. First, the part wedges tailored to the surface with motorized means to
geometry is estimated by the first echo time-of-arrival follow-
ing one of several possible strategies: pulse–echo, pitch–catch,
correct small misalignments have been used [1], [2].
or plane wave. These are analyzed with regard to their per- In many situations, however, the probe-part geometry
formances and acquisition time, giving closed formulae to get is not sufficiently well known to accurately compute the
the coordinates of interface points. After a curve fitting and focal laws. This is a common case when the manufac-
extrapolation process, a virtual array that operates in a homo- turing process has no tight tolerances (casting material,
geneous medium is computed, avoiding the complications of
refraction at the interface and allowing operation with already
molded components, non-rigid parts or parts with vary-
known focusing hardware. This hardware is initialized with the ing shape, soft curved plates, weld-caps, etc.) or when
set of focusing parameters adapted to the estimated probe– the relative positions of the probe and the part change.
part geometry, and ensures that all received samples are in Merely a tenth of a millimeter of displacement yields tim-
focus. Using a standard computer, the auto-focusing procedure ing errors of 130 ns in water immersion, which is more
currently takes about 2 s to perform. Experiments carried out
under different conditions validate the proposed technique.
than half the period of a 5 MHz signal; a misalignment
of the transducer probe by only 1° may lead to amplitude
losses of several decibels and incorrect localization of the
I. Introduction indications by some millimeters [3].
Water-filled bags attached to conventional array probes

P hased-array technology applied to nondestructive can be useful to adapt the probe to irregularities in the in-
testing (NDT) provides high-resolution images gen- spected part [4]. However, refraction effects must be again
erated by scanning an inspected part with electronical- taken into account, unless the propagation velocities of
ly controlled ultrasonic beams. Dynamic depth focusing the two media are similar (i.e., biological tissue). Another
provides higher sensitivity and quality by setting a focus approach uses flexible array probes, in which individual
at every image sample. However, unlike optical cameras, elements are spring loaded, forcing them to operate in
their acoustic counterparts lack the auto-focusing feature. contact [5], [6]. Methods to estimate the position of the ar-
The availability of such a function in NDT instruments is ray elements with auxiliary devices or from image features
a dream held by many engineers and operators in the field. have been developed, showing good behavior, although
Automatic dynamic depth focusing (ADDF) would avoid with high computational cost [7]–[9]. Furthermore, these
the complications of computing and setting the correct fo- methods demand low-speed scanning to avoid excessive
cal laws, while maintaining the best possible resolution at wear of the transducer surface and use nonconventional
all depths, without needing computer-aided design (CAD) probes and equipment.
geometry descriptions or additional software tools. The time reversal mirror technique provides automatic
Currently, planar parts can be inspected with the array focusing to the strongest reflector [10], [11]. Generaliza-
probe in contact. This is the simplest scenario, in which tion to multiple reflector detection has been proposed us-
the focal laws can be computed using closed formulae from ing iterative techniques [12], [13]. These methods demand
the knowledge of the array geometry and the sound propa- specialized hardware and have high processing loads, so
gation velocity. Furthermore, dynamic depth focusing fo- that their applications have been mostly constrained to
the academic field. Furthermore, the state-of-the-art of
these methods still requires the geometry description to
form an image.
Manuscript received November 2, 2013; accepted January 8, 2014. Recently, the self-adaptive ultrasound technique
This work has been funded by project DPI 2010-17648 of the Spanish (SAUL) has been proposed as an auto-focusing alternative
Ministry for Science and Innovation. [14], [15]. Starting with a simultaneous trigger of all the
J. Camacho, J. F. Cruza, and C. Fritsch are with the Ultrasound Sys-
tems and Technology Group (USTG), Spanish National Research Coun- array elements, the two-way time-of-flight to individual
cil (CSIC), Madrid, Spain (e-mail: [email protected]). elements is used to set the emission focal laws for the next
J. Brizuela is with the Comisión Nacional de Energía Atómica (CNEA), trigger event. After some shots, a wave-front parallel to
Consejo Nacional de Investigaciones Científicas y Técnicas (CONICET),
Buenos Aires, Argentina. the interface is produced and the ultrasound penetrates
DOI http://dx.doi.org/10.1109/TUFFC.2014.2955 with normal incidence. The technique yields unfocused

0885–3010 © 2014 IEEE

674 IEEE Transactions on Ultrasonics, Ferroelectrics, and Frequency Control , vol. 61, no. 4, April 2014

Fig. 1. Sector imaging of an arbitrarily-shaped part with a coupling

medium.

images that are useful to inspect laminated structures, Fig. 2. Detected interface points using the pulse–echo technique.
and requires a full-parallel ultrasound system with a high
channel count to get adequate coverage of the interface.
In addition to the accurate knowledge of the array–part Special attention is paid to interface geometry estima-
geometry, refraction at the interface complicates the focal tion, which is a subject poorly covered in ultrasound lit-
law computation task. Unlike the homogeneous case with erature. Although this topic was previously addressed for
probe and part in contact, there are no closed formulae pulsed radar [30] and ultrasonic range-finders [31], appli-
to compute the time-of-flight from every array element to cation of these methods in NDT ultrasound is not entirely
every focus. This problem is frequently addressed by simu- evident. In this work, three techniques are proposed and
lation tools, using different approaches: spatial impulse analyzed: pulse–echo, pitch–catch, and plane wave. They
response [16], [17], propagation of monochromatic waves are also experimentally tested and compared. It is shown
[18]–[20], ray-tracing methods [21]–[24], etc. In general, that the proposed auto-focusing procedure provides good
computing time can be an issue for dynamic depth focus- geometry estimation and image quality comparable to
ing, so that focal laws are frequently obtained for a small that obtained by conventional methods that require an a
set of foci [25]. priori and accurate knowledge of the array–part geometry.
From a practical point of view, Fermat’s principle is
frequently applied in searching for the minimum time-of-
flight path. In the most general case, without code opti- II. Interface Geometry Estimation
mization, the required fine sampling of the interface leads
to long searches. In [26], using the total focusing method The inspection of arbitrarily shaped parts is frequently
(TFM), focal law computation took 49 s per image with performed inserting a coupling material (liquid or solid)
parallel computing on GPU/CPUs. For a phased array, between the array and the part (Fig. 1) and generating a
we proposed a fast focal law calculator (FFLC) based on linear or sector scan. When the geometry is unknown, it
the Newton–Raphson algorithm [27], which takes 2.6 s for must be first estimated. Here, three techniques are pro-
an image of similar size in a standard computer. Real- posed to perform this task using first echo time-of-arrival
time processing has been recently achieved for TFM using measurements: pulse–echo, pitch–catch, and plane wave.
a top-performance GPU platform, but for small images
(30 × 10 mm) and relatively low element count (32 array A. Pulse–Echo Interface Estimation
elements) [28]; the processing time is proportional to the
number of image pixels and to the number of array ele- The interface profile can be obtained with N individual
ments. shots by measuring the time-of-flight in pulse–echo from
This work addresses auto-focusing in isotropic media every array element to the interface, Ti, 1 ≤ i ≤ N. Fig.
with soft, curved interfaces of arbitrary shape, using stan- 2 shows the points detected in a cylindrical part of radius
dard array probes and equipment. For this purpose, the RC = 32 mm, in water immersion with a coupling path
array–part geometry is estimated by first echo time-of- of 15 mm, by an N = 64-element array with d = 0.5 mm
arrival measurements and curve fitting algorithms. Two pitch (sound propagation velocity c = 1480 m/s). Only
more steps follow, computing a virtual array and setting the fraction of the interface where there is normal inci-
the parameters of a real-time dynamic focusing hardware, dence from the array elements is directly detected.
as is described in our recent work [29]. Of course, any fo- The distance from array element i to the interface is Ri
cal-law computing procedure can be used once the probe– = cTi/2. This yields a set of N circumferences centered
part geometry has been determined, which is the major at the array elements and tangent to the interface, as is
concern of this paper. shown in Fig. 3. Array element i has coordinates (xAi, zAi),
camacho et al.: automatic dynamic depth focusing for ndt 675

Fig. 3. The interface is defined by the envelope of N circumferences.

Fig. 4. Detected interface points using three trigger events and multiple
receivers in pitch–catch.
the element pitch is d, and the angle to the tangency point
is given by:
time TABik from element i to the interface and back to
R − R i +1 c element k are, respectively:
sin φi = i = (T − Ti +1). (1)
d 2d i
TAi = 2R Ai /c, TABik = (R1ik + R 2ik )/c . (3)
Because Ri is normal to the interface, the coordinates
of P are As before, the measured time TAi determines a circum-
ference of radius RAi, whereas the pitch–catch value TABik
x Pi = x Ai + Ri sin φi defines an ellipse with foci Fi and Fk at elements i and k,
(2)
z Pi = z Ai − Ri cos φi . respectively, for which

The only assumption made is that segments Ri and R Ai = cTAi /2, R ABik = R1ik + R 2ik = cTABik . (4)
Ri+1 are parallel. As long as d is small in relation with the
local curvature radius of the interface, errors produced by Assuming that the normal to the interface has approxi-
this approximation will be low. In fact, for a plane inter- mately equal slope at P and at E, the angle αik of the
face (RC = ∞), this method is free from geometric errors. incident ray with the normal at P is equal to the angle of
The technique uses N trigger events to get N − 1 inter- the reflected ray at E. Therefore, triangles PEFi and PEC
face points. The acquisition time can be reduced by using are equal and PC = PFi = RAi and EC = EFi = R1ik. Now,
one out of M < N elements in pulse–echo, which results in from the cosine theorem applied to FiCFk,
~N/M trigger events instead of N with d in (1) changed to
M · d. For example, for N = 64, M = 8 yields 7 interface
points {(xPi, zPi)} with 8 measurements. However, reduc-
ing the acquisition time this way has some effects. The
larger the distance M · d is, the greater are the differences
in the interface slope at P and at Q, the assumption of
parallelism between Ri and Ri+M is not as accurate, and
geometric errors increase.

B. Pitch–Catch Interface Estimation

In this case, element i is used as emitter–receiver in

pulse–echo and a set of M elements at every side are used
as receivers in pitch–catch configuration. Fig. 4 shows the
same example as in Fig. 2 for pitch–catch mode with M =
10 and emitters at elements {i} = {11, 32, 53}. In this case,
60 interface points are detected with just 3 shots.
Fig. 5 shows the geometry for elements i (emitter) and
k (receiver) in the pitch–catch mode. The emitter-to-re-
ceiver distance is Dik = (k − i) · d. The pulse–echo time-
of-flight TAi to element i at (xAi, zAi) and the pitch–catch Fig. 5. Geometry of the interface profile estimation by pitch–catch.
676 IEEE Transactions on Ultrasonics, Ferroelectrics, and Frequency Control , vol. 61, no. 4, April 2014

Dik2 = 4R A2 i + (R1ik + R 2ik )2 − 4R Ai(R1ik + R 2ik ) cos αik

(5)
4R A2 i + R AB
2 2
ik − D ik
cos αik = . (6)
4R AiR ABik

Substitution in RAi = R1ik cos αik yields

4R A2 iR ABik
R1ik = , (7)
4R A2 i + R AB2 2
ik − D ik

where all the terms at the right-hand side are known.

Then,

R 2ik = R ABik − R1ik . (8)

Fig. 6. Detected interface points with plane-wave excitation.
The segments gik and hik in Fig. 5 verify

h ik2 = R12ik − g ik2 = R 22ik − (Dik − g ik )2 (9) sion, all the elements are used as receivers and the individ-
R12ik − R 22ik + Dik2 ual first echo arrival times are measured. Fig. 6 shows an
g ik = , h ik = R12ik − g ik2 , (10) example of this process, were only part of the plane wave
2Dik
produces reflections at the interface that arrive at some
and the coordinates (xEik, zEik) of the interface at Eik are array element. Fig. 7 shows the basic geometry for this
acquisition mode. Element i receives the echo from point
x Eik = x Ai + g ik, z Eik = z Ai − h ik . (11) P at time Ti after the wave has traveled the distances zi
and Ri. Taking into account that zi < 0,
In this way, for every trigger shot with 2M receivers
Ri − z i = cTi . (12)
surrounding the emitter, a set of 2M interface points
E(xEik, zEik) is obtained. Although it would also be pos-
sible to compute RBi = RAiR2i/R1i because of the similar- The locus of possible points P(x, z) is given by
ity of triangles FiPE and FkQE, the errors for estimating
(x − x Ai )2 + z 2 − z = cTi (13)
Q would be rather high because of the departure of the
interface from the straight line passing by P and E, so this (x − x Ai )2 cTi
z = − . (14)
option is not advisable. 2cTi 2
With M = 1, two interface points are obtained for every
three elements, using the central element as the emitter. This parabola with focus at the receiving element
In this way, ~N/3 trigger events produce ~2 · N/3 inter- (xAi, 0) is tangent to the interface at P, so that their slopes
face points. The geometric errors will be lower than those are equal:
found by the pulse–echo technique because the assump-
tion of parallelism of the normal at P and E fits better dz x − x Ai
= = tan αi . (15)
than at P and Q. dx cTi
In an extreme case, this technique would yield N − 2
interface points with a single trigger from a central ele-
ment, with M = N/2 − 1. However, care must be taken
because geometric errors increase with the distance Dik
from the emitter to the receivers. Also, the element sen-
sitivity decreases with higher steering angles, making the
method less robust because weak interface echoes can be
masked by noise. As with the pulse–echo mode, there is
a trade-off between interface estimation accuracy and ac-
quisition time.

C. Plane Wave Interface Estimation

In this method, a plane wave is produced, up to some

range within the near field, by simultaneously triggering
all the array elements (zero-delay focal law). After emis- Fig. 7. Geometry for the interface estimation with a plane wave.
camacho et al.: automatic dynamic depth focusing for ndt 677

From Snell’s law at P, the incident and reflected angles After estimation of S interface points, a curve fitting al-
are equal to αi. Then, gorithm interpolates and extends the interface beyond the
detected region. The robustness of this approach depends
z i = −Ri cos 2αi on the measurements’ accuracy, the absence of outliers,
(16)
x i = x Ai + Ri sin 2αi . and, to a lesser extent, on the number S of points, as is
discussed later.
Substitution into (12) yields When the interface has a completely unknown shape,
but soft curvature, a second- to fourth-degree polynomial
cTi fit can be used. Abrupt changes or discontinuities would
Ri = . (17)
1 + cos 2αi require specific fitting algorithms, like piecewise approxi-
mations. If the interface is known to be planar, a fit to a
Calling ΔLi = c(Ti+1 − Ti) and assuming that normal first-degree function is advisable. Finally, if the interface
at P and Q are parallel: is circular, the modified least square algorithm (MLS) de-
scribed in [32] is quite effective. These procedures are very
c(Ti +1 − Ti ) ∆Li fast in modern computers and yield an analytical repre-
R i +1 − R i = = . (18)
1 + cos 2αi 1 + cos 2αi sentation of the interface shape.
The geometric errors associated with these techniques
From (16), are mainly due to the assumption of parallelism among
the normals at different points of the interface and will de-
− cos 2αi
∆z i = z i +1 − z i = ∆Li crease for higher interface curvature radius. Fig. 8 shows
1 + cos 2αi
(19) the maximum absolute geometric errors as a function of
sin 2αi the normalized curvature radius RC/DA with DA = N ·
∆x i = x i +1 − x i = d + ∆Li.
1 + cos 2αi d for the three described techniques and the considered
example (N = 64, d = 0.6 mm, water immersion, and cy-
After a few mathematical manipulations, the following lindrical parts at 15 mm from the array). No other sources
equations are obtained: of error are considered here. The estimated interface is
extrapolated using the MLS method to get enough points
∆z i (tan 2 αi − 1)∆Li for inspection at normal incidence.
tan αi = = (20)
∆x i 2(d + ∆Li tan αi ) The higher errors correspond to the plane wave tech-
∆Li tan 2 αi + 2d tan αi + ∆Li = 0. (21) nique, although these are quite acceptable for most appli-
cations (below 2 μm for a curvature radius RC = 0.5DA,
16 mm in this case). The lowest errors correspond to the
Solving for tan αi and substituting into (15) yields x =
pitch–catch technique, where the detected interface points
xi, whereas z = zi is obtained from (14), giving the coor-
get closer. It is worth noting that, in all cases, the geomet-
dinates (xi, zi) of P. Applying this procedure to all the ele-
ric errors become negligible from a practical point of view.
ments 1 ≤ i ≤ N provides a set of N − 1 interface points
Geometric errors introduce some deviation in the cur-
with a single simultaneous trigger event.
vature radius estimation, shown in Fig. 9 for the three

III. Comparison of the Proposed Techniques

The three methods proposed to estimate the interface

geometry have different characteristics. The pulse–echo
technique is the simplest one and can be used with stan-
dard multiplexed phased-array equipment, where a linear
scan with an active aperture of 1 element is performed
and the first echo arrival-time is recorded. By contrast,
the pitch–catch and plane wave techniques require parallel
reception with a high element count, which needs higher
cost full-parallel systems. Processing time is not an issue
for any of these techniques, because they are based on
closed formulae.
On the other hand, for a given number S of estimated
interface points, the plane wave technique requires a single
simultaneous trigger event, being the faster one. By con-
trast, the pulse–echo and the pitch–catch methods will re-
quire several trigger events, typically around S/5 to S/10, Fig. 8. Maximum geometric errors of the proposed techniques as a func-
choosing M ~ 5. tion of the normalized curvature radius of the inspected part.
678 IEEE Transactions on Ultrasonics, Ferroelectrics, and Frequency Control , vol. 61, no. 4, April 2014

Fig. 9. Absolute errors in the estimation of the curvature radius. Fig. 11. Geometric errors with a 160 MHz sampling rate and water im-
mersion.

proposed methods. It is higher with the plane wave meth-

od and is negligible with the pulse–echo and pitch–catch measurements can be expected. They may become slightly
techniques. reduced by the curve fitting process as long as the number
Every technique estimates a different extent of the in- p of free parameters is well below the number S of data
terface. Fig. 10 shows, in degrees, the angular region of the points acquired. This is another reason to use procedures
interface covered by the different techniques as a function that detect a larger number of interface points. Moreover,
of the relative curvature radius. The plane wave procedure sampling errors can be further reduced using correlation
yields a smaller range (less than 25° versus more than 40° or frequency-domain methods to measure the time-of-ar-
for the lower radius part), which makes the fitting algo- rival of echoes.
rithm less robust. These differences decrease for higher Fig. 11 shows the effect of time quantization using a
curvature radius. 160 MHz sampling rate. Geometric errors (around 10 µm)
As has been shown, the intrinsic geometric errors of the are well above those found before with infinite time resolu-
proposed techniques are so low that they can be neglected tion (intrinsic geometric errors), but are quite acceptable
in practice. However, in real applications, there are other for estimating the interface shape from a practical point
sources of error that may be much more significant. of view.
First, the time-of-flight is quantized by the sampling A different and more dangerous error source is the pres-
rate, which introduces an uncertainty of at least ±1 sam- ence of outliers or time measurements that do not belong
pling period. For example, in water immersion and with a to echoes produced by the interface. Because only the time
sampling rate of 160 MHz, errors of about ±5 µm in the to the first echo is recorded, outliers can be produced by
the excitation tail, noise spikes, or weak interface echoes.
The estimated interface may be seriously affected by outli-
ers, which should be removed.
Defining a blind region near the array surface will avoid
detecting the excitation pulse tail as an echo. Also, keep-
ing only the time-of-flight measurements that lie within
a given time window around their neighbors, is an effec-
tive way to remove outliers and spurious measurements.
Other methods reject those measurements that are some
distance apart from the estimated geometry, which can be
iteratively evaluated.
Another source of error may be the part geometry it-
self. The proposed techniques are quite robust for detect-
ing convex shapes, even with relatively small curvature
radius. However concave parts must have a large enough
curvature radius, especially with the plane wave technique,
because of the focusing effect of the interface toward the
array center. For concave parts of small curvature radius,
Fig. 10. Interface extent in degrees of the estimated points by the differ- the crossing of the returning rays invalidates the plane-
ent procedures. wave interface estimation procedure. By contrast, the
camacho et al.: automatic dynamic depth focusing for ndt 679

pulse–echo and pitch–catch techniques are more robust the corresponding real element A to focus F following the
and allow estimating the geometry of concave parts as refraction laws through the entry-point at G. Mathemati-
well. cally,
Finally, further errors can be introduced by uncertain-
ties on the input parameters. Although the array pitch is tVF + t K ≈ t AF . (22)
given with high precision by the manufacturer, sound ve-
locity in the first medium is prone to errors resulting from By application of the Abbe’s invariant,
temperature variations and material composition. To min-
imize their impact, it is advisable to calibrate the system  c2 
measuring the sound velocity before interface detection. t K = t AG  1 − 12  , (23)
 c2 

where tAG is the time-of-flight from A to G, obtained along

IV. Application to Automatic Focusing with the time-of-flight tAFB from element A to focus FB at
(xFB, zFB) by numerical methods (applying Fermat’s prin-
After estimation of the interface geometry, focal laws ciple, for example). The time-of-flight tAFA from element A
must be computed and programmed to obtain the image. at (xA, zA) to focus FA at (xFA, zFA) is also computed and
As proposed in [29], this is achieved in two steps: first, the coordinates (xV, zV) of the virtual element at V are
the complications of refraction are avoided by obtaining obtained by solving the equations
a nearly equivalent virtual array that operates in a single
medium (the second one); then, a real-time focusing hard- (xV − x FA)2 + (zV − z FA)2
ware per element is configured with just a few parameters. + t K = t AFA
c2
Because these procedures were covered in depth in the (24)
referred work, only a short presentation will be given here (xV − x FB )2 + (zV − z FB )2
+ t K = t AFB .
in relation to their application to automatic focusing. c2

A. Obtaining a Virtual Array Operating The virtual array provides exact time-of-flight to foci
in a Single Medium FA and FB and approximate values to points in between.
Properly choosing FA and FB yields a good approximation
Focal law computation becomes complicated by refrac- of (22) in the FA to FB range. As was shown, the small
tion at the interface when there are two media with differ- errors involved in this approximation (a few nanoseconds)
ent propagation velocities, c1 and c2. The evaluation of an have little impact on NDT images.
equivalent virtual array that operates only in the second
medium completely removes this problem (Fig. 12). B. Using a Real-Time Dynamic Depth Focusing Hardware
For this purpose, it is postulated that the time-of-flight
tVF of echoes from a virtual array element V to a focus F Over the years, different approaches have been pro-
following a straight path at propagation velocity c2, plus posed to perform real-time DDF for homogeneous media
a constant tK, is approximately equal to the time tAF from using specialized hardware [33]–[38]. In general, the focus-
ing circuits iteratively compute the time-of-flight, the fo-
cusing delays, or the sampling instants to consecutive foci
from a starting point. These circuits require setting some
parameters before operation, which are easily computed
from known variables (steering angle, element position,
start sample, etc.).
However, these approaches do not work when changes
of propagation media are involved. However, once the in-
terface has been removed by the virtual array, most of
these circuits can be used to perform real-time DDF. In
particular, the architecture proposed in [29] offers several
advantages for auto-focusing: it requires initialization of
only 2 parameters, the array elements can be located any-
where (as required by the virtual array) and it consumes
few hardware resources. Furthermore, it provides a timing
resolution of TS /v, where TS is the sampling period and
v is an arbitrary parameter, keeping focusing delay errors
within ±TS /v at all depths. For a virtual array of size
Fig. 12. Geometry of the virtual array that operates in the second me- DV, the minimum range of operation with full aperture is
dium. given by
680 IEEE Transactions on Ultrasonics, Ferroelectrics, and Frequency Control , vol. 61, no. 4, April 2014

v the actual array–part geometry and without further op-

R0 = D . (25) erator intervention.
4 v −1 V

For shorter ranges, dynamic aperture can be used. Op-

eration with numerical apertures of f-number ≥ 1 for val- V. Experimental Verification
ues v ≤ 15 is possible.
The focusing circuit requires initialization of only two Several experiments have been carried out to assess the
registers, whose values are computed using closed formu- performance of the proposed auto-focusing approach. Al-
lae from the virtual element position, the steering angle, though all of them were carried out in water immersion
and the sampling period. Every element requires its own for convenience, it is important to note that the method
focusing circuit, but its implementation with the DSP can be also applied with solid wedges. The first experi-
cells available in state-of-the art FPGAs allows time-mul- ment compares the three proposed methods (pulse–echo,
tiplexing of several channels with a single cell. pitch–catch, and plane wave) for estimation of the inter-
face geometry.
A 5-MHz, N = 128-elements array (Imasonic, Besan-
C. Autofocus Operation
çon, France) with pitch d = 0.6 mm was used to inspect,
by water immersion, a 50-mm-radius aluminum ring lo-
Auto-focusing is a very valuable tool to set up inspec- cated at approximately 34 mm from the array surface. A
tions, because it provides a simple and automatic means SITAU-112 ultrasound system (Dasel S.L., Madrid, Spain)
to free the operator from setting focal laws or concern with 128 active channels was used to generate and receive
about CAD-based geometry descriptions. In the current the ultrasound signals, controlled by a script developed
non-optimized experimental implementation, it takes in Matlab. Signals sampled at 40 MHz were upsampled
about 2 s to perform the whole process using a standard by a factor of 4, giving an equivalent sampling rate of
computer and a Matlab (The MathWorks Inc., Natick, 160 MHz. All processing was carried out on an Asus K53S
MA) program. When the operator pushes the autofocus (AusTek Computer Inc., Fremont, CA) portable computer
button, the following process is run: with an Intel Core i7 (Intel Corp., Santa Clara, CA) 2.2-
GHz processor and 4 GB of RAM memory.
1) Estimate the interface geometry using the pulse– In all experiments, time-of-flight measurements were
echo, pitch–catch, or plane-wave methods as de- carried out from the trigger instant to the maximum of
scribed in Section II. The acquisition time depends the envelope by programming a hardware gate with a
on the method used. threshold to avoid detecting noise. Despite emitting with
2) For every scan-line, compute the corresponding vir- a single element in pulse–echo and pitch–catch modes, a
tual array. Currently, this task consumes about 3/4 moderate gain was enough and noise was not an issue in
of the time used by the auto-focusing function. either experiment because the coupling medium has low
3) Compute the focusing parameters and program them attenuation.
in the hardware, which is fast enough to be neglected Fig. 13(left) shows the points detected by the pulse–
with regard to the previous tasks. echo, pitch–catch, and plane-wave interface estimation
methods using 128 array elements, together with the fit-
Upon completion of these steps, the hardware is ready ted circles. Maximum part coverage (±24°) is obtained
to acquire images with DDF, automatically adapted to with the pulse–echo technique, whereas pitch–catch and

Fig. 13. (left) Estimated interface points (dots) and adjusted circle with (top) pulse–echo method, (center) pitch–catch with M = 15 and (bottom)
plane-wave emission. (right) Reconstruction errors.
camacho et al.: automatic dynamic depth focusing for ndt 681

TABLE I. Interface Geometry Estimation Results.

Radius Water column Coverage RMS reconstruction
Method (mm) (mm) (°) error (µm)
Pulse–echo 49.67 33.81 ±24 7
Pitch–catch 49.84 33.80 ±22 13
Plane-wave 49.85 33.80 ±16 16

plane-wave methods obtain ±22° and ±16° respectively. In A second experiment was conducted using an alumi-
the pitch–catch method (left, center) it can be appreciated num 90° sector located some unknown distance apart from
how the detected points become grouped around the four the array probe, which is arbitrarily tilted with regard to
emitting elements. the part surface [Fig. 15(a)]. After performing the auto-
Reconstruction errors are evaluated as the distances focusing process fitting of the interface to a third-degree
of the detected points to the estimated interface. They polynomial, a sector image was obtained using an active
are comparable for all methods and do not show any pat- aperture of 96 elements [Fig. 15(b)]. Fig. 15(c) shows a
tern or tendency along the array elements (Fig. 13, right), detail around the two groups of SDHs. The effect of DDF
which confirms a good fitting to a circle. RMS errors were keeping a good lateral resolution and sensitivity at all
7, 13, and 16 μm for the pulse–echo, pitch–catch, and depths can be appreciated.
plane-wave methods, respectively, being of the same order In a third experiment, an arbitrarily shaped concave
as those found by simulation (Fig. 11). aluminum part with several SDHs was used [Fig. 16(a)].
Table I summarizes the results for the three acquisition The water path from the center of the array probe was ap-
methods. All of them estimate the part radius and the proximately 35 mm. Fig. 16(b) shows the B-scan obtained
distance to the array with good precision, the maximum by pulse–echo from every array element, with the straight
difference being 180 μm and 10 μm, respectively. line showing the end of the blind region used to suppress
Fig. 14(left) shows a picture of the experimental ar- the pulser tail and marks where echoes were detected. Fig.
rangement; Fig. 14 (right) shows the image obtained by 16(d) shows the calculated interface points and the ex-
linear scanning with 48 active elements and normal inci- trapolated interface, fitting to a third-degree polynomial
dence after applying the proposed auto-focusing method by the least squares method.
(pulse–echo estimation of the interface was used). The After estimating the interface, virtual array and focus-
interface, the bottom echo, and the 1.5-mm side-drilled ing parameters were computed for a sector scan of ± 45°
holes (SDHs) are correctly imaged and with the expected using 128 active elements. The resulting phased-array im-
lateral resolution according to the aperture size. The am- age is shown in Fig. 16(c). All the SDHs in the field of
plitude of the leftmost SDH in the second row is lower view are clearly seen, with the expected lateral resolution
because of shadowing effects. All images in this section are according to the aperture size. A slight improvement on
represented with a linear color scale. the resolution, barely perceptible in the image, was ob-
Images obtained with pitch–catch and plane-wave de- tained by fitting to a sixth-degree instead of a third-degree
tection methods are not shown because they are indistin- polynomial.
guishable from that obtained with the pulse–echo tech- Fig. 17(a) shows a final example with the same part
nique, as was expected because of the small differences in but in a region where the surface changes its slope. The
the interface estimation (Table I). From here to the end interface estimated by fitting a third-degree polynomial is
of the document, the detection method used is pulse–echo shown in Fig. 17(b). Note that although it does not ex-
with 128 elements. actly match the part shape at the left, the detected region

Fig. 14. (left) Experimental arrangement, aluminum ring with 3 rows of pairs of SDHs. (right) Linear scan image following the auto-focusing process
with the pulse–echo technique.
682 IEEE Transactions on Ultrasonics, Ferroelectrics, and Frequency Control , vol. 61, no. 4, April 2014

Fig. 15. (a) Probe-part arrangement; (b) phased array sector image; (c) detail of the SDHs.

Fig. 16. (a) Inspection of a concave part in water immersion; (b) linear scan with 1 active element to detect the interface; (c) obtained image with
128 active elements and ± 45° steering; (d) representation of the array probe and the calculated interface.

Fig. 17. (a) Probe-part geometry; (b) detected points and extrapolated interface; and (c) linear scan image.
camacho et al.: automatic dynamic depth focusing for ndt 683

extent is enough to perform a linear scan at 0° and 32 ac- References

tive elements, as shown in Fig. 17(c).
[1] F. Young, C. Uyehara, and H. Bui, “Ultrasonic testing of corner ra-
dii having different angles and sizes,” U.S. Patent US2009/02111361
A1, Aug. 27, 2009.
VI. Conclusions [2] R. Meier, J. Becker, and T. Rehfeldt, “Method for the ultrasound
testing of a workpiece within a curved region of its surface and de-
vice suitable for the execution of the process,” U.S. Patent 7 516 664
This work proposes an automatic dynamic focusing B2, Apr. 14, 2009.
[3] S. Chatillon, S. Mahaut, and P. Dubois, “Simulation of advanced
procedure particularly addressed to applications in which UT phased array techniques with matrix probes and dynamic set-
the probe–part geometry is unknown. The auto-focusing tings for complex component inspections,” Rev. Quant. Nondestruct.
function requires knowledge of the propagation velocities Eval., vol. 28, pp. 864–871, 2009.
[4] R. Long, J. Rusell, and P. Cawley, “Non-destructive inspection of
in both media and the element pitch of the array. No components with irregular surfaces using a conformable ultrasonic
further information is required. It produces dynamical- phased array,” in 6th Int. Conf. NDE in Relation to Structural Integ-
ly depth focused images of high quality, competing with rity for Nuclear and Pressurized Components, 2007, pp. 1–9.
[5] O. Casula, C. Poidevin, G. Cattiaux, and G. Fleury, “A flexible
those acquired with focal laws computed from a perfectly phased array transducer for contact examination of components
known geometry using specialized software tools. Further- with complex geometry,” in 16th Conf. Nondestructive Testing, 2004,
more, it can be applied with conventional array probes art. no. 773.
[6] G. Toullelan, A. Nadim, O. Casula, P. Dumas, E. Abittan, and L.
and equipment. Doudet, “Inspection of complex geometries using flexible phased ar-
Auto-focusing avoids the burden of computing and pro- ray transducers,” in 17th World Conf. Nondestructive Testing, 2008,
gramming focal laws and the complications of CAD-based art. no. 391.
[7] O. Roy, S. Mahaut, and S. Chatillon, “Ultrasonic inspection of com-
geometry descriptions, and is an effective tool to adapt to plex geometry component specimen with a smart flexible contact
changes in the probe–part geometry during the inspection. phased array transducer: Modeling and applications,” in Proc. IEEE
It is even faster than conventional procedures based on Ultrasonics Symp., 2000, pp. 763–766.
[8] A. Hunter, B. Drinkwater, and P. Wilcox, “Autofocusing ultrasonic
software tools to compute the focal laws: the current non- imagery for non-destructive testing and evaluation of specimens
optimized implementation (Matlab code), computes and with complicated geometries,” NDT Int., vol. 43, no. 2, pp. 78–85,
programs the required dynamic depth focusing informa- 2010.
[9] A. Hunter, B. Drinkwater, and P. Wilcox, “Least-squares estimation
tion in about 2 s using a standard PC for a 128-element of imaging parameters for an ultrasonic array using known geomet-
array and 100 steering angles. ric image features,” IEEE Trans. Ultrason. Ferroelectr. Freq. Control,
The auto-focusing procedure is based on three steps: vol. 58, no. 2, pp. 414–426, 2011.
[10] M. Fink, C. Prada, F. Wu, and D. Cassereau, “Self focusing in in-
1) estimation of the array-part geometry, 2) computing a homogeneous media with ‘time reversal’ acoustic mirrors,” in Proc.
virtual array for every steering angle, and 3) computing IEEE Ultrasonics Symp., 1989, pp. 681–686.
the small set of parameters needed by a real-time dynamic [11] M. Fink, “Time reversal of ultrasonic fields—Part I: Basic prin-
ciples,” IEEE Trans. Ultrason. Ferroelectr. Freq. Control, vol. 39, no.
focusing circuit. In this work, three methods have been 5, pp. 555–566, 1992.
analyzed to perform the first step, based on pulse–echo, [12] C. Prada, F. Wu, and M. Fink, “The iterative time reversal mirror:
pitch–catch, and plane wave measurements of the first A solution to self-focusing in pulse–echo mode,” J. Acoust. Soc. Am.,
vol. 90, no. 2, pp. 1119–1129, 1991.
echo time-of-arrival. [13] C. Prada, M. Tanter, and M. Fink, “Flaw detection in solid with
Theoretically and experimentally, it has been shown the DORT method,” in Proc. IEEE Ultrasonics Symp., 1997, pp.
that, with regard to accuracy, all of these methods are 679–683.
[14] S. Robert, O. Casula, O. Roy, and G. Neau, “Real-time inspection of
equivalent for the estimation of the interface geometry. In complex composite structures with a self-adaptive ultrasonic tech-
practice, their main differences reside in the time spent to nique,” in 18th World Conf. Non Destructive Testing, 2012, art. no.
perform the measurements for a given number of detected Tu.4.B.2.
[15] S. Robert, O. Casula, M. Njiki, and O. Roy, “Assessment of real-
interface points. The pulse–echo method is the slower one, time techniques for ultrasonic non- destructive testing,” Rev. Quant.
requiring N measurements to get N − 1 interface points. Nondestruct. Eval., vol. 31, pp. 1960–1967, 2012.
By contrast, the plane wave technique obtains N − 1 in- [16] P. R. Stepanishen, “Transient radiation from pistons in an infinite
planar baffle,” J. Acoust. Soc. Am., vol. 49, no. 5B, pp. 1629–1638,
terface points with a single simultaneous emission and the 1971.
pitch–catch one gets about the same number of points [17] F. Buiochi, O. Martínez, L. G. Ullate, and F. Montero, “A compu-
with a few shots. However, although the pulse–echo tech- tational method to calculate the longitudinal wave evolution caused
by interfaces between isotropic media,” IEEE Trans. Ultrason. Fer-
nique can be applied with standard phased array equip- roelectr. Freq. Control, vol. 51, no. 2, pp. 181–192, 2004.
ment, the pitch–catch and the plane wave methods require [18] T. Huttunen, M. Malinen, J. P. Kaipio, P. J. White, and K. Hynynen,
full-parallel receiving instruments. “A full-wave Helmholz model for continuous-wave ultrasound trans-
mission,” IEEE Trans. Ultrason. Ferroelectr. Freq. Control, vol. 52,
Experimental work is in agreement with theoretical no. 3, pp. 397–409, 2005.
predicted errors, which are of a few micrometers in the [19] A. Ibañez, M. Parrilla and J. Villazón, “Simulation of ultrasonic
estimation of curvature radius and position. These errors continuous wave fields in homogeneous media with soft curved in-
terfaces,” in 9th European Conf. NDT, 2006, art. no. Th.3.3.1.
may be even lower than those produced in real applica- [20] A. Ibañez, C. Fritsch, M. Parrilla, and J. Villazón, “Monochromatic
tions using CAD descriptions of the geometry because of transfer matrix method for acoustic field simulation thorough media
undetected small misalignments between the probe and boundaries,” Phys. Proc., vol. 3, no. 1, pp. 883–890, 2010.
[21] L. Leber, O. Roy, and N. Jazayeri, “Applications of phased array
the part, whereas auto-focusing considers their actual true techniques to NDT of industrial structures,” in 2nd Int. Conf. Tech-
relative positions. nical Inspection and NDT, 2008, art. no. 72.
684 IEEE Transactions on Ultrasonics, Ferroelectrics, and Frequency Control , vol. 61, no. 4, April 2014

[22] D. Richard, D. Reilly, J. Berlanger, and G. Maes, (2010, Sep.). Jorge F. Cruza was born in Madrid, Spain, in
“Software tools for the design of phased array UT inspection tech- 1983. He obtained the B.S. degree in telecommu-
niques,” in Simulation in NDT 2010, [Online]. Available: http:// nication engineering in 2009 and the M.S. degree
www.ndt.net/article/SimNDT2010/papers/9_Richard.pdf in electronics in 2011, both from the Universidad
[23] C. Poidevin, O. Roy, S. Chatillon, and G. Cattiaux, “Simulation de Alcalá (UAH). He is currently Ph.D. student at
tools for ultrasonic testing,” in Proc. 3rd Int. Conf. NDE in Relation UAH. Since 2010, he has been an Associate Re-
to Structural Integrity for Nuclear and Pressurized Component, 2001, searcher at the Spanish National Research Coun-
art. no. 105. cil (CSIC). His research interests include high-
[24] P. Calmon, S. Mahaut, S. Chatillon, and O. Roy, “Simulation of resolution acoustic imaging, real-time
phased array techniques for realistic NDE configurations,” Proc. beamforming, and FPGA signal processing.
IEEE Ultrasonic Symp., 2002, pp. 723–727.
[25] J. Liang, Z. Wang, Y. Shi, “Ultrasonic inspection of thick parts with
phased array dynamic focusing,” in European Conf. NDT 2010, art.
no. 1.3.91. Jose Brizuela was born in Córdoba, Argentina,
[26] M. Suttcliffe, M. Weston, B. Dutton, and I. Cooper, “Real-time in 1977. He received the B.S. degree in electronic
full matrix capture with auto-focusing of known geometry through engineering from the National Technological Uni-
dual layered media,” in NDT Conf. of the British Institute of Non- versity of Argentina (UTN-FRC) in 2002. Then,
Destructive Testing, 2012, art. no. 3A1. in 2007, he obtained the M.S. degree in informa-
[27] M. Parrilla, J. Brizuela, J. Camacho, A. Ibaez, P. Nevado, and C. tion technology from the Technical University of
Fritsch, “Dynamic focusing through arbitrary geometry interfaces,” Madrid, Spain (EUI-UPM). In 2010, he achieved
Proc. IEEE Ultrasonics Symp., 2008, pp. 1195–1198. his doctoral degree from the Complutense Univer-
[28] M. Sutcliffe, M. Weston, P. Charlton, K. Donne, B. Wright, and I. sity of Madrid, Spain (UCM). Since 2005, he has
Cooper, “Full matrix capture with time-efficient auto-focusing of been working on research activities at the Spanish
unknown geometry through dual-layered media,” Insight, vol. 55, National Research Council (CSIC). In 2012, he
no. 6, pp. 297–301, Jun. 2013. moved to Argentina to work as a research scientist for the National Sci-
[29] J. F. Cruza, J. Camacho, L. Serrano-Iribarnegaray, and C. Frit- entific and Technical Research Council (CONICET) and technology con-
sch, “New method for real-time focusing through interfaces,” IEEE sultant for the National Atomic Energy Commission (CNEA). His re-
Trans. Ultrason. Ferroelectr. Freq. Control, vol. 60, no. 4, pp. 739– search interests include nondestructive testing (NDT) by ultrasound
751, 2013. techniques (UT) for industrial and medical applications, ultrasonic imag-
[30] T. Sakamoto and T. Sato, “A target shape estimation algorithm ing, beamforming, phased array systems, conventional UT and advanced
for pulse radar systems based on boundary scattering transform,” backscattering UT (AUBT) techniques, embedded and real time sys-
IEICE Trans. Comm. E, vol. 87-B, no. 5, pp. 1357–1365, 2004. tems, FPGA & GPU programming, data processing, and system integra-
[31] B. Barshan, “Ultrasonic surface profile estimation by spatial vot- tion and transfer technology activities.
ing,” in Proc. IEEE Instrum. Meas. Technology Conf., 2001, pp.
583–588.
[32] D. Umbach and K. N. Jones, “A few methods for fitting circles to
data,” IEEE Trans. Instrum. Meas., vol. 52, no. 6, pp. 1881–1885, Carlos Fritsch was born in Madrid, Spain, in
2003. 1948. He obtained the M.S. degree in telecommu-
[33] K. Jeon, M. H. Bae, S. B. Park, and S. D. Kim, “An efficient real nication engineering and the Ph.D. degree in in-
time focusing delay calculation in ultrasonic imaging systems,” Ul- formatics from the Universidad Politécnica de
trason. Imaging, vol. 16, no. 4, pp. 231–248, 1994. Madrid, in 1971 and 1988, respectively. Since
[34] M. Bae, “Focusing delay calculation method for real-time digital 1972, he has been a member of the scientific staff
focusing and apparatus adopting the same,” U.S. Patent 5 836 881, at the Spanish National Research Council (CSIC).
1998. In the last 22 years, he has headed several re-
[35] H. T. Feldkämper, R. Schwann, V. Gierenz, and T. G. Noll, “Low search projects in the field of ultrasound technol-
power delay calculation for digital beamforming in handheld ultra- ogy. His research interests include digital signal
sound systems,” in Proc. IEEE Ultrasonics Symp., 2000, vol. 2, pp. processing applied to the field of ultrasound,
1763–1766. beamforming and processing architectures, ultrasound imaging, fast fo-
[36] B. G. Tomov and J. A. Jensen, “Compact implementation of dy- cal law computing algorithms, and new methods for NDT and medical
namic receive apodization in ultrasound scanners,” Proc. SPIE, vol. applications.
5373, pp. 260–271, 2004.
[37] I. Lie and M. E. Tanase, “About the possibility to implement a
nonuniform oversampling receive beamformer in a FPGA,” in Proc.
IEEE Ultrasonics Symp., 2005, pp. 1404–1407.
[38] R. Alexandru, “Delay controller for ultrasound receive beamformer,”
U.S. Patent 7 804 736 B2, 2010.

Jorge Camacho was born in Montevideo, Uru-

guay, in 1979. He received the B.S degree in elec-
tronic engineering from the Universidad de la
República, Uruguay (UdeLaR), in 2004. He re-
ceived the Ph.D. degree in systems and automa-
tion engineering from the Universidad Com-
plutense de Madrid (UCM) in 2010. Since 2005,
he has been an associate researcher at the Spanish
National Research Council (CSIC). His research
interests include high-resolution acoustic imaging,
real-time beamforming, and new imaging methods
for medical and NDT applications.

View publication stats