Int. J. Pres. Ves.

& Piping 20 (1985) 17-41

Design of a Real-Time Inspection System for NDE of

Reactor Vessels and Piping Components

S. G a n a p a t h y & B. S c h m u l t

Ultrasonic Imaging Laboratory, Department of Electrical and Computer Engineering,

The University of Michigan, Ann Arbor, MI 48109, USA

(Received: 18 March, 1985)

ABSTRACT

Progress in the development of a special-purpose system for use in a real

time in-service inspection system for reactor vessels and piping
components is described in this report. An analysis of the synthetic
aperture processing algorithm is presented and new methods of speedup
are described. A number of special purpose processor architectures are
presented and two of the more promising ones are described in detail and
are compared and evaluated. Proposed specifications for an initial field
inspection system are presented. A brief description of the capabilities of
a laboratory prototype processor (to be fabricated) is given.

INTRODUCTION

It has been apparent for some time that more advanced techniques are
required for reliable in-service inspection of nuclear reactor pressure
vessels and piping components. Several such advanced systems are under
development, including the A L O K system, 1 the Acoustic Holography
system 2 and systems based on Synthetic Aperture Focusing Techniques
or SAFT. It has been reported previously 3 that a SAFT-based system can
give an order of magnitude improvement in resolution over conventional
inspection methods.
17
18 S. Ganapathy, B. Schrnult

The Ultrasonic Imaging Laboratory at the University of Michigan has

been developing a SAFT-based inspection system with two goals in mind.
The first goal is to develop hardware and algorithms that can handle the
computational complexities of real-time SAFT, since previous implemen-
tations of SAFT were found to be inadequate for reactor inspection. 4
Secondly, we have been developing designs for a complete system capable
of real-time in-service inspection in the field. This paper describes results
of work in both of these areas.
Any system suitable for use in the field must be able to inspect three
major classes of reactor components:
1. Circumferential and longitudinal welds on the pressure vessel
walls.
2. All welds on piping coming from the vessel.
3. Nozzles, where pipes join the vessel wall.
It is expected that a SAFT-based system will be able to handle all three of
these areas. However, the complexity of these parts varies and nozzles are
particularly difficult. The main objective of this research is to create an
initial field system that will prove the viability of SAFT inspection
techniques. For this reason the system described in this paper has been
kept simple and will only inspect the first class of components (vessel
welds).

SCANNING l
PROCESSING

Fig. 1.
t!- DISPLAY

Different phases of SAFT-UT.

DECISION
MAKING j

A SAFT inspection system as envisaged by us consists of four major

parts: Scanning, Processing, Display and Automated Decision Making
(see Fig. 1). Each of these areas will be described in detail in the next four
sections. This paper concludes an initial design of a complete system for
field inspection and presents current plans for the fabrication of
prototypes.
Documentation on all aspects of research done by this laboratory is
available in the form of reports submitted to the United States Nuclear
Regulatory Commission (hereinafter the NRC). The interested reader is
referred to Refs. 3, 5-10.
 Real-time inspection of reactor vessels and piping components 19

SCANNERS F O R SAFT-BASED SYSTEMS

The scanning system deserves attention here because the scanning

methods represent a departure from normal practices. Data collection is
done on an A-scan basis with each A-scan representing amplitudes going
down into the wall thickness. Note that most of the SAFT system is
digital. As soon as an A-scan is collected it is digitized and the digital data
are passed to the rest of the system.
Our laboratory scanning is currently done with focused transducers
moving in a point to point fashion, using a raster scan pattern as shown in
Fig. 2. The distance between any two adjacent points will be equal to the
wavelength of the current transducer. The sampling rate, which
determines how precisely measurements are taken in the depth direction,
is expected to correspond to a distance equal to one eighth of the
wavelength. At 2.25 MHz Operation the wavelength will be about 0" 1 in.

)×
Fig. 2. Scanning in a raster pattern.

The field system must use the same type of raster scan pattern as is used
in the laboratory. This is necessary for two reasons. First, the collection
points must be aligned on a grid for accurate processing. If the scanner is
allowed to go around the entire reactor, the points on the second row will
not be aligned as well as they would if they had been collected on a raster
pattern, since the transducer covers a much smaller total distance to fill an
aperture while raster scanning. Secondly, collected data must be stored
until a complete two-dimensional aperture is filled. If an entire
20 S. Ganapathy, B. Schmult

circumference were collected at a time, many such datasets would have to

be collected in order to get one full aperture. This would require a very
large storage area for raw data. Raster scanning requires much less.
The scanning mechanism must also be capable of maintaining a
constant angle of incidence with the front surface and capable of
maintaining the transducer at a constant distance from the surface. The
system will permit both normal and squint mode scanning, with shear
waves being included in squint scanning.
Since the definition of real-time operation is to process the data as fast
as they become available, the maximum speed of the scanner determines
how fast the rest of the system must operate. Currently available
mechanisms can operate at about 0.1 in per 0" 1 s in point to point mode,
with signal averaging. The maximum size of an A-scan is expected to be
about 12 in. Thus on a 0.1 × 0" 1 in grid, this amounts to a collection rate
of 1.2 in3/s or 72 in3/min. Since it is possible for speedup to come from
mechanical improvements or by use of a linear array of transducers, the
processor is being designed to handle collection rates of 250 300 ina/min.
Note, however, that some of the collected data are required only to fill out
apertures of center A-scans on the boundary of the area being imaged and
will not themselves be processed. For typical scanning tasks these could
account for around 25 % of the total data, so the data rate in the rest of the
system need only be around 200--250 in3/min.
Scanners suitable for field use are currently under development by
other NRC contractors.

D I G I T A L SYNTHETIC A P E R T U R E P R O C E S S I N G

The computer processing time involved in SAFT has long been

considered to be too high for use in the field. The development of
techniques to do SAFT processing in real time has been one of the major
accomplishments of this laboratory. The research to attain this goal has
been done in three phases. First, concepts and algorithms were developed
to determine exactly what processing had to be done. Secondly, the
algorithms were studied and benchmarks and simulations were done to
determine how the processing could be speeded up. Finally, based on the
results of this analysis, special purpose hardware has been designed that
will handle the real-time data rates. The rest of this section describes the
results from these three areas.
 Real-time inspection of reactor vessels and piping components 21

Concepts of digital SAFF processing

The purpose of SAFT processing is to increase the lateral resolution of

ultrasonic images. 5 The processing of a single target (center point) is
illustrated in Fig. 3. The set of points around the center A-scan that
receive a response from the center point constitute the aperture. To
process the center point it must be averaged with one contributing point
from each A-scan in the aperture. The time shift for a particular off-center
A-scan is the distance down to the point that contains the response from
the center point, which is equal to the distance from the center point to the
surface at the site of the off-center A-scan. The shift is represented in
Fig. 3 as a broken line. To SAFT-process an entire data volume, this
process of averaging all points over an aperture must be repeated for each
point in the volume.
Thus SAFT processing consists of a number of executions of a basic
cycle. This cycle involves these four steps:

1. locate the next off-center A-scan;

2. calculate the time shift;
3. fetch the point;
4. add the point to the center point.

These steps vary in complexity. The first step is very simple, consisting of a
table look-up to the coordinates of the next A-scan. Fetching the
contributing point is also a memory reference. Both steps can be speeded
up through the use of high-speed memory technology. (Note, however,
that the memory reference in step 3 may be to a mass memory that is by
nature much slower than the memory used to store A-scan coordinates,
and this is a potential bottleneck.) The addition can be done very quickly
by using high-speed logic and well-known addition techniques. The shift
calculation is then the most complex step of the four and will be given
additional attention below.
The shift calculation for the general case is displayed in Fig. 4. The
solid curve is an actual arbitrary surface. The dashed line represents an
ideal machined surface that is used for the software machining technique
to be described later. The vertical lines represent A-scans. The shift is the
distance from the center point to the surface at the position of the off-
center A-scan:
shift = [(d + h) 2 + r2] 1/2
 Y

Z' : :

CENTER
A-SCAN
OFF-CENTER
A-SCAN
Fig. 3. Relationship between center point, off-center A-scan, and time shift in SAFI
processing.

~/ '6
h

Fig. 4. SAFT time shift calculation for general case.

 Real-time inspection of reactor vessels and piping components 23

where d is the depth of the center point from the surface directly over it, h
is the difference in surface heights between the center and off-center A-
scans, and r is the radial distance from the center A-scan to the off-center
A-scan.
Although the equation is mathematically simple, the multiplications
and particularly the square root are time-consuming operations. These
operations make the shift calculation the most time-consuming step in the
basic cycle. The shift calculation can be speeded up substantially just by
eliminating the square root operation. This is possible under certain
conditions.
If the object being scanned has a surface with the proper symmetry then
shift calculations are greatly simplified. This will be illustrated with a fiat
sample although the same simplification holds for other types of
geometries. For further details on the use of look-up tables for other
geometries see Appendix A of Ref. 6. For the case shown in Fig. 5:
shift = (d 2 + r2) 1/2
Although this case looks very similar to the previous general case, there is
an important difference. The two shift parameters, d and r, are now
independent of the position of the center A-scan. So, if adequate memory
is available, the shifts for an A-scan can be computed once, and then used
for every A-scan that is to be processed. For this technique, the
parameters d and r would be used to index into the table. This table look-
up is much faster than the multiplications and square root operations
used in the general case.
This look-up table technique can actually be done with a one-
dimensional look-up table, instead of the two-dimensional one described
r

Fig. 5. SAFT time shift calculation for a flat surface.

24 S. Ganapathy, B. Schmult

above. Such a table is much smaller and can be accessed faster. Since
many center points can be imaged at a single depth, a look-up table can be
calculated for that depth only. When the depth changes a new table is
calculated or read in from secondary memory. Note that this read or
calculation could go on while the previous depth is being processed.
In summary, a one-dimensional look-up table does not require much
memory and will produce time shifts in around 30 ns, which is much faster
than calculating them with square roots. The ability to do this results
from the invariance of shifts with respect to the (x,y) position of the
center A-scan, which is a direct result of the symmetrical surface.

Computational analysis of SAFT

The performance of a SAFT processor will be determined by the time
taken to execute the basic cycle or inner loop. Before a processor can be
designed the maximum allowable cycle time must be known. This section
analyzes the SAFT algorithm and determines the number of basic
operations per unit volume that must be performed as a function of
processing parameters such as frequency and the included angle of the
aperture. When these figures are combined with performance require-
ments the maximum basic cycle time can be determined.
The most complex part of the basic operation is the time delay
calculation which involves path lengths and sound velocities in both the
sample and the coupling medium. The path length calculation requires
the coordinates of the transducer at each position and, depending on how
the transducer is focused, the calculation could be a simple direct
computation or a more complicated iterative process. Algorithms for
calculating time delays for flat and cylindrical surfaces for various cases of
focusing at the surface and above and below the front surface are derived
and can be found in earlier reports. 5
The overall speed of the synthetic aperture processing is also dependent
on how much is known about the boundary and the geometry of the
synthetic aperture. The speed of the process is bounded by the following
two extreme cases:
1. Nothing is known about the sampling boundary except that the
transducer has been made to track it and record its coordinates at
each scan point. This means that the entire set of time delay
calculations must be repeated for every point in every A-scan every
time another off-axis A-scan within the aperture is to be used.
 Real-time inspection of reactor w'ssels and piping components 25

. The sampling boundary is mathematically describable and the

aperture geometry remains constant over the entire boundary.
The most common examples are a plane, a cylinder, or a sphere.
Since the aperture geometry remains constant, the time delay
calculations for the entire aperture need to be done only once.
They could then be stored in a look-up table and used to process
the entire sample.

Nevertheless, real-time synthetic aperture processing remains a formid-

able task even for the second case.
This subsection will analyze the computational complexity of
expanding square aperture SAFT. Although real apertures will
probably not be square, their size will be close to that of a square
aperture and it is the number of points accessed that determines
processing time.
Consider a point A in a metal block (Fig. 6) of thickness d and let the
included angle of the aperture be 2tp. Then the width of the aperture
required at A is 2ztan q~. To achieve good resolution with SAFT, the
sampling interval in the X and Y dimensions should be at most one
wavelength (one 2). Hence the number of points in the square aperture is
'z tan ~o'~
~--W-) + ly (1)

which is
4(zt )2
Therefore the number of additions to the point A is

4 ( ~ ) 2 + 4 ( z 2tan q~) (2)

An A-scan of N z points corresponds to a depth of

z = 0.5V6N~

where 6=sampling interval, V=velocity of sound in metal,

N z = number of points in A-scan with depth z, and 0-5 = allowance for
round trip of sound.
The sampling interval in the Z dimension must be consistent with
sampling theory to avoid aliasing. This means that the sampling
 ~
26 S. Ganapathy, B. Schmult

nsducer

z d

Fig. 6. Metal block of depth d.

frequency must be at least twice the frequency of the transducer and

preferably twice the upper frequency in the spectrum of the transducer.
Thus
1
rf
wherefis the center frequency of the transducer and r is a constant chosen
such that usually 2 < r < 8.
Therefore,
2rz
Nz- t
or
Nz).
2r
For z = d:
2rd
Na- i
Therefore, the total number of computations needed per A-scan for a
depth of d is given by substituting eqn (2) into eqn (1) to yield
2rd/2
N~
4 2 I ( ~ tan q~)2 + (~-r tan q~)l
Nz=l
 Real-time inspection of reactor vessels and piping components 27

The total number of computations T needed for a block of x y d is given

approximately by
2rd/2
Nz xy
4 ~-~ I(~-r tan q~)2 + (~-t~tan ~0)] 22
Nz=l

The total number of computations per unit volume is given approxi-

mately by
T
xyd
which gives
2rd/2
4 Nz N~

Nz= 1

Similarly, the amount of memory required to contain a single aperture

block could be derived from eqn (1) to be
2rd/2

~-~ I2 (2~tan q~) + 112 (4)

Nz=l

The complexity per unit volume of SAFT vs. depth as computed from
eqn (3) has been plotted in Fig. 7. As can be seen, the center frequency of
the beam as well as the included angle have a large impact on the
computational complexity involved. In fact, the complexity is of order
(a2
~-~tan 2q~).
,
Memory requirement is also an important parameter to be considered
in a SAFT processor design. Figure 8 (calculated from eqn (4)) is a plot of
memory size needed to contain an aperture for different depths d.
Although sophisticated memory management could cut the amount of
memory needed, Fig. 8 is nevertheless a good indication of how different
parameters could affect the memory requirements.
From the Figures, it can be concluded that the processor speed and
memory size requirements are highly dependent on the depth, the beam
frequency and the included angle. Decreasing any of the above
parameters would reduce the requirements substantially. However, the
beam frequency and the included angle are also parameters that affect the
28 S. Ganapathy, B. Schmult

Degree Resolution
..0 30 ° 0'0Z
- - + 1 MHZ
..... & 2 MHz ..........rn..... .,....n 50" 0.03
109 .... • 3 MHz .D....aa ..~ ........El"" A 40 ° 0.04
....... O 4 MHz d~ PdPIB
............. 0 5 MHz
0.... .,'D'' . . O ' " " ~ 60" 0'03
......':" .,:, .........
D" o""~ ....o":~6" ~, :j~:-:'" ~ '
....... •...... . ""~' ~."-"::'~ ~ 40" 0"05
....~ ..,-,...::...o- ..j~.::.~'- .o.....o.--B
•
'~"~:" :
" ' ~r "' ~"""
" ~Cr"
~ :~.< ~ ' ~ -~ll 5 ~ 0{)4
0"06
(~05
108
.... .,,-'- _ . 40"

'" ," /~-r ..I ~I" ~-I30 ° 0"08

/'/" JIl~'" I " .~ I I -- ~I'"
,/ .~, _~..., ..i i~.~ ....~ 60" 0-06
m • .~,/ #~ /~ i ~ ~ ..., ~-"
c
.9 .," j . • I / 11 ~.,./" ~ . ~ . ur
. . . . .'" ~ r 0'08
• I / / _ 1 ~" .~.~
107
/ P •' / / i / • 1/I / . ~ " / ~ ¢ . - ' &-.- _ t ' _ . ~ 40" 0'10
/ . / ~, /. .~"
I I" I / /~.i. ~. j~,.t ~.,I.~ 30" 0'12
0
// // ./'~:./-~" ~./'~ ~..~.~'~
e~

d / /
/ ,~ . / / ~ . / . ~ .~./'~
Z '/•./~," ,~" ,~.-"
106 / ~././-/ //~
/ ./)~' ,,4' ~ 60" O-lZ
/ I . ~ / ~,,/

a// / ~ ~p,1-- ~ 40" 049

•• 30" 0.25

105

104
1 2 3 4 5 6 ? 8 9 10 11 12
Thickness, d (in)
F i g . 7. Operations per unit volume as a function of depth. N = (8rd 2 tan 2 ~b)/325.
 ... - :..~ ' \ ,.~', \ , ,~ ,.. \\ \ \ \
:-. :-. \ '.. . -5\ ~ \ ', ' \ \ ",. \ '\ \
•: .~
- ~- '\ -:, :
\ "-.A~
':A'\ \
~ ,\ ~
\ ".
\. ~ \
\"~ \
"\ \
"\
0
=:
".. , . "- , '-X\ \ \ \ \ \ \ \ \
•... ":: \ . . : \ '...\\. \.., \ \ \ .\ '\ \ 0
(I:) ~'
[:2 o , ~ q ~ UQ =m .~. U ~ h
~..~,~/.~..~".:'. '" '. \%,,~ ~, e- e" ~
•-... -..... \ .-....-.\ -...~\ \\\. \ ., ' .\ \
'.. ". " ".. "- '..X'. \ " \ \. \\\', "
~\~ ~, <
•.. .........\ .......-\. '.~.:,\. \,',,:\\".\",,\.\ ".\•\ "\
"c .......•~ ......""-0.-o•.\'...\
'-a.o '.~"....~",,::o,.,...\• ,~,:,,
.,',,.,, \ \
... "... . . .....~.. \\ "- ~ ~-- -\ .\
0
......r-, 'E]."O.."O.'b. '~J::). m.o."m. ~'. t '~.\ ~,
,.\ '~ \
N N N N N "... ~... ~ . "~ \
Z
+ ~moB
...-L
kl I I I I I Jill i i t I I ~lll I I I I [ II ~ I I i I I
(O~O)~'--~l~r,,k ") -4" ~ N ~e..,O")(O~(Ol.¢ ') "d" ~ ~OO')'00~'.,-E) t~ -4' o
P
X
seanlaedols),ujod ~o 'ON N
30 S. Ganapathy, B. Schmult

resolution. The resolution of SAFT is of the order of 2/(4 sin~0). Higher

frequencies or higher included angles give better resolution but also result
in higher computational costs. A good compromise between com-
putational requirement (cost) and resolution is being analyzed and will be
reported in the future.
Once the resolution is fixed, it can be achieved in most cases by
increasing the included angle or increasing the center frequency of the
transducer. Because computation only goes up as the square of the
included angle (for small angles tan ~p = ~p) and goes up as the fifth power
of frequency (both approximations), it is always better to increase the
angle whenever it is appropriate and feasible. As an example, consider the
two points marked A (5 MHz, 40 °) and B (3 MHz, 60 °) in Fig. 7. While
both combinations of parameters give a resolution of 0.04in, the
combination at point A using 5 MHz requires five times the number of
computations as does point B.

Speedup attempts for SAFT

We have implemented several SAFT processing systems in software on

our laboratory computer systems. These programs have given best-case
processing rates of around 2in3/min for laboratory conditions. As
pointed out above, a field system will require processing rates of the order
of 200 in3/min. A basic question then is whether a conventional computer
can handle the processing rates for field conditions. From the results of
the previous section one can determine how fast a SAFT processor must
operate (in terms of the execution time of the basic cycle) under given
conditions. This section summarizes work done to determine how fast
conventional processors can operate under any conditions.
The results of this work, reported more fully in a recent report, 9
involved benchmarks done on several available computers and show that
a special-purpose system is needed to achieve the required speeds. The
measure used in these benchmarks was the time to shift and add one off-
center data point to a center point, referred to here as z (see Fig. 3 to
visualize the operation). This value was determined for several computers
through program benchmarks or simulations and the results are given in
Table 1. Note that the model of SAFT computation used to compute r
assumed the use of look-up tables. If this assumption does not hold for
use in the field then these z values will be increased substantially. Also
included is the corresponding processing rate for a 12 in thick sample.
 Real-time inspection o f reactor vessels and piping components 31

This assumes a frequency of 2.25 MHz and an included aperture angle of

45 °
From Table 1, it can be seen that while available minicomputers, like
the VAX or PRIME, may give a reasonable speed for small volumes, the
processing rate is not sufficient for real-time synthetic aperture processing
under field conditions. The FPS-120B array processor shows consider-
able improvement over the minicomputer and the cost is reasonable, but
much of the processor is wasted. The CRAY-1 was used as a boundary
condition and shows a significant improvement in speed, but the cost is
several million dollars.

TABLE 1
SAFT Inner Loop Processing Time for Various Processors

Processor r (ps) In3/min based on Approximate

12 in thick sample cost

lnterdata 7/32 14.750 0" 18 $50 000

Perkin-Elmer 3240 6.940 0.38 $110 000
D.E.C. VAX-I 1,/780 5.500 0.47 $140000
Prime 750 4.500 b 0.58 $150 000
F PS- 120B" 0.500 5.22 $60 000
CRAY- I 0.100 26.10 $8 000 000
Special purpose
(I 6 processor) 0.010 261 $150 000

" FPS-120B is an Array Processor that must be attached to a mainframe.

b Estimated.

The algorithms used in these benchmarks were all implemented in soft-

ware, but could be used just as effectively in a special-purpose hardware
system. With a given system the goal is to reduce the number of operations
needed to shift and add one off-center data point to a center data point.
This is done by eliminating memory references and performing operations
in parallel. Any of the four steps in the basic cycle (see the section on
concepts of digital SAFT processing) could consist of several machine
instructions, depending on the application and assumptions made.
Therefore, all four operations must be considered when developing a
system. The programs we used to benchmark used a table look-up for the
shift value and thus the shift calculation assumed a symmetric surface. In
general, the shift calculation could be much more complicated, since it
involves determining the distance between two points in a three-
32 S. Ganapathy, B. Schmult

dimensional coordinate system. This involves multiplications and square

roots, which are much more time consuming than an add or m e m o r y
reference. It is then desirable to be able to use a table look-up for the shift
calculation to reduce the required processing time.
F r o m these benchmarks it is evident that currently available computers
do not offer the required performance at a reasonable cost and thus a
special-purpose processor is the most viable alternative. This would allow
parallelism to be exploited to best advantage, while only including the
necessary hardware. The design of such a processor is the subject of the
next section.

A SAFT processor architecture

The results of the previous section make it apparent that a parallel

architecture is required to achieve real-time speeds. Such a parallel
processor puts n identical processing elements (PEs) to work at the same
time and gives a speedup of n', n ' < n. Data will be processed n ~ times as
fast as if a single processor were used, or alternatively, if a total cycle time of
-c seconds is required, each PE can be operated with a cycle time of n'z
seconds, allowing slower hardware to get the job done.
Several different parallel architectures have been considered and one,
known as the data broadcast architecture, has been selected for
implementation. 1o The application of parallelism to SAFT processing by
this architecture is illustrated in Fig. 9. There are two PEs assigned to two
adjacent center A-scans. (In general, n PEs will be assigned to n adjacent
center A-scans.) The surrounding A-scans represent the union of the
apertures of both center A-scans. In operation each point from each A-
scan is broadcast to both PEs. Each PE decides where (if anywhere) each
particular data point contributes to its center A-scan. After all A-scans
have been broadcast, there will be two completely processed center A-
scans.
Most of the points broadcast are in the apertures of both PEs. Points
on the left and right boundaries are only in the aperture of one, so when
these points are broadcast one of the PEs will be idle. This means that
using n PEs will not get data processed n times as fast as using just one.
However, the speedup is sufficiently high for real-time operation. We
expect that a PE can be built to operate with a cycle time of about 160 ns.
For apertures 63 points wide, a SAFT processor with 10 PEs will offer a
speedup of about 8.75, resulting in an overall cycle time of about 18 ns.
 Real-time inspection oJ reactor vessels and piping components 33

FE2 APERTURE
I
I

FEI PE2

!
t .1
I
PEI APERTURE
Fig. 9. Combined aperture forseveralprocessing elements.

This is under the 20 ns figure given previously as a requirement for field

operation.
The organization of the PEs is illustrated in Fig. 10. Each PE has
storage for its center A-scan and a queue for incoming data points. The
basis for shift calculations and aperture membership determination is a
set of control tables. The table determines how an off-center A-scan must
be shifted and added to a center A-scan. There is one table for each A-scan
position in the aperture and one c o m m a n d in the table for each point in an
A-scan.
As an A-scan is broadcast the appropriate control table is sent, one
c o m m a n d at a time, to each PE. Each c o m m a n d can be A D D , S H I F T or
A D D & SHIFT. Each c o m m a n d is executed by its PE as it arrives. After
one A-scan is broadcast, the next A-scan will be the next in a line parallel
to the PE line. Thus the table that one PE needs for the next A-scan will be
the same table used by its neighbor on the previous A-scan. So an external
controller feeds a new table to one PE and each PE loads its used
commands into a buffer. The other end of each buffer feeds the next PE. A
more detailed explanation of this operation is given in section 3 of Ref. 10.
The simplicity and lack of computation in the PEs make it easy to build
high-speed PEs with standard hardware. The major reason this speed is
possible is the use of table look-up methods. This means all shift
34 S. Ganapathy, B. Schmult

BROADCAST
DATA

t DATA
C~UEUE

CENTER

CONTROL ADDER
A-SCAN

T
POIHTER
--~ READOUT

PREVIOUS CONTROL NEXT

ELEMENT > ELEMENT
DELAY

Fig. 10, Block diagram of a processing element.

calculations are done once and incorporated into the control tables.
Compensation for surface variations is obtained with the software
machining technique.I° Because raw A-scans are machined before being
written into main memory, the PEs can implement pure table look-up
SAFT, giving high-speed operation with tolerance for surface variations.

Design of a SAFT processor

Figure 11 shows a block diagram of a complete SAFT processor. This
processor forms a part of an overall inspection system. It is connected to
 PROCESSOR EL£M~NT
ARRAY

%
I
L ...
5
rI[l~ lib SOF'*I'WAR ~
%
A--SCANS I-t-NTERF-ACE I'IAIN HE'rtO~Y 2"
HA CH]'t4 ~i~G ~NT£RF'Ar'~

F r
I COLCR CCD~$
I
C~£D~ DA~A
LOOK--UP t 2
CON'IROLLER C~RE£N) ADDR~ AND SH~F~"
TABL~
CBLU~) CONTROL

XO S~C
Fig. 11. Block diagram of a SAFT processor.
36 S. Ganapathy, B. Schmult

and controlled by a minicomputer that forms the core of the whole

system.
The controller coordinates all operations in the processor based on
commands received from the minicomputer. Raw data and surface
position information arrive directly from the scanner system. The
software machining module adjusts the A-scans to compensate for
surface variation, and the updated A-scans are written into main memory.
The controller reads A-scans out of main memory as needed and
broadcasts them to the processor element array, along with the control
tables. Control tables, like other parameters, are calculated by the
minicomputer and loaded before processing begins. When complete
processed A-scans are available they are read out of the PE array and
output, usually to an envelope detector.

DISPLAY T E C H N I Q U E S

The SAFT-based inspection system incorporates several new display

techniques that attempt to display images of flaw regions rather than raw
A-scan data. The principal display method will be a technique called real-
time walkthru. The details of this technique appear in a NRC report
currently in preparation.
The basis for walkthru is pseudo color imaging. This technique is used
on two-dimensional planes of data that are cross-sections of the collected
data volume. Each data point in the plane is assigned a color based on its
relative amplitude. The color scale assigns 'cold' colors like blue to low
background amplitudes and 'hot' colors such as red to peaks.
The walkthru technique presents data as a sequence of pseudo colored
planes. Each plane represents a single row of A-scans and is a cross-section
of the weld perpendicular to the weld centerline. New planes will overwrite
old ones as they become available and thus give the effect of ~walking
through' the weld region. Updating the display device for a new plane
should require only about 64K bytes of data and several seconds are
available to transmit it (the time to scan one row). Thus real-time
operation will be easy to achieve and the display can be done as the data
are collected.
There will also be available several forms of ~non-real-time' display
techniques. As data are collected they are not only displayed but also
stored on a disk system. These other display methods operate on these
 Real-time #tspection of reactor vessels and piping components 37

stored data. The first technique is an interactive walkthru. The

operator can inspect any desired region and can modify the display rate to
a maximum of about two frames per second. Planes can also be frozen, or
the sequence reversed. If the scan takes a long time this technique can
also be used to view the data after they have been collected, in much less
time than the scan takes. It should be noted that a laboratory version of
walkthru has already been implemented, completely in software. Under a
timesharing system it gives speeds of around 2.25-3 s per frame. Thus a
speedup factor of only about 5 is necessary to achieve the two frame per
second goal. This should be easily accomplished with hardware aids.
Specific planes or regions located with walkthru can then be subject to
additional display techniques. Some of these are perspective contour
plots and display of shaded images to convey 3-D orientation. Perspective
views can be generated from any user viewing angle and further it is
possible to rotate these images to convey complete three-dimensional
information about depth and orientation of flaws inside an inspection
volume. Hardcopy can also be made of single displays. Finally, it is
possible to retrieve saved data from previous scans to rerun the displays.

A U T O M A T E D DECISION M A K I N G

Automated decision making is the least developed of the four phases

outlined in the introduction. Research into the basic problems in this area
is under way. The basis for decision making is the extraction of the
boundary of the flaw to convey size and shape information. Two- and
three-dimensional edge detectors are being developed that will perform
this extraction.
It is not expected that automated decision making will be incorporated
into an initial field system.

T O W A R D A C O M P L E T E I N S P E C T I O N SYSTEM

The four phases described in the introduction must be combined to form a

complete inspection system. An initial fieldable system can be put
together with only the first three phases (scanning, processing and
display). A block diagram of such a system appears in Fig. 12.
The entire system is built around and controlled by a general purpose
COLOR CODES
(BLACK} GENERAL ~NTE~CON~JECT ENVIFLOPE SCAN
<RED) A--~CAN PAPA SQANNER
DETECTOR J ~ l pRG~ESSOR
~.~AP-T JIC C~I'-ITROL
~LtJE) ATTACHED pI~JDCE~OH
CONTROL

l I
~J
c~

..~YSTFM CONTROL
< >
COt.IPU'T ER

I
t SYSTEM DATA "r~PE
DTSKS D~SK~5 ~:rvkGs
[ 1
AL -r EK rJ~.'TE COLO~ ~ D'rSpLAY
£~'SPLA'( j DISPLAY
P~.OCE~SCR
DEVICE

Fig. 12. Block diagram of initial SAFT-based inspection system.

 Real-time inspection of reactor vessels and piping components 39

16 or 32 bit minicomputer known as the system control computer (SCC).

It manages all the collected data and uses external devices and processors
to complete functions. All operator interaction takes place through
terminals connected to the SCC.
The scanner is the only unit actually attached to the reactor. It consists
of a multiple axis mechanism, data collection electronics and a
microcomputer controller. It can receive direct positioning commands
from the SCC and return individual A-scans for analysis. In normal
operation it is given a region to scan by the SCC and will do the scanning
without intervention. Surface mapping information, if kept, is sent to the
SCC for storage. Raw data and surface height are sent directly to the
SAFT processor.
The SAFT processor also runs without intervention. It simply receives
raw A-scans and outputs processed A-scans at the same rate. All
processing parameters, such as A-scan size, are loaded from the SCC
before processing starts. The shift table and aperture profile definition are
both encompassed by the control tables. These tables are computed once
by the SCC and are also preloaded.
Processed data are still RF and are unsuitable for display and thus
envelope detected. Real-time rates require special-purpose hardware or at
least a dedicated high-speed microprocessor. Like the SAFT processor it
just sits in the data stream and outputs A-scans at the same rate as they
come in.
The disk system is used to store programs for the SCC and data for the
current weld. The current weld data can also be saved on magnetic tape
for future use. Thus it is possible to go back and review previous scans, or
to compare two scans of the same flaw to check for growth.
Note that this system organization permits easy expansion. Future
complex operations can be added simply by building a special-purpose
box that does the operation and inserting it into the data stream, just like
the SAFT processor. Processing support for automated decision making
can be implemented in this manner.

CONCLUSIONS

This paper has presented the results of research toward the design of a
complete SAFT-based inspection system for field use. The area thought
to be the major bottleneck, SAFT processing, has been overcome and the
40 s. Ganapathy, B. Schmult

fabrication of a prototype processor to demonstrate this is under way.

Likewise, we feel that a display system can be fabricated that will display
data at the same rate as they are collected. These two components will
then be able to keep up with collection rates of current and future
scanners, thus constituting a real-time system. Scanners suitable for use
for this system are currently under development by other N R C
contractors and should be available in the near future. Completion of this
final development and fabrication will permit the construction of an initial
fieldable system that can demonstrate the capabilities of SAFT. The
future addition of automated decision making and other enhancements is
expected to create an advanced N D E system uniquely suited for the
inspection of nuclear reactor vessels and piping components.

ACKNOWLEDGEMENTS

This research has been supported in part by the Nuclear Regulatory

Commission under contract NRC-05-75-182. Thanks are also due to
several members of the Ultrasonic Imaging Laboratory for contributions
to several aspects of the system design.

REFERENCES

1. Grohs, B. et al., Characterization of flaw location, shape, and dimensions

with the ALOK system, Materials Evaluation, 40 (1982), pp. 84 9.
2. Kiefer, R., Schmitz, V. and Wendel, H., Fast recording and numerical
reconstruction of holograms, Proc. 3rd Int. Conf on Nondestructive
Evaluation in the Nuclear Industry, Salt Lake City, Utah, 11 13 February,
1980, pp. 499-512.
3. Frederick, J. R., Seydel, J. A. and Fairchild, R. C., Improved ultrasonic
nondestructive testing of pressure vessels first annual report, Nuclear
Regulatory Commission Report NUREG-O007-1, January 1976.
4. Ganapathy, S., Wu, W. S. and Schmult, B. Analysis and design
considerations for a real-time system for nondestrctive evaluation in the
nuclear industry, to appear in Ultrasonics.
5. Frederick, J. R., Fairchild, R. C. and Anderson, B. H., Improved ultrasonic
nondestructive testing of pressure vessels, Nuclear Regulator), Commission
Report NUREG-O007-2, September 1977.
6. Frederick, J. R., Vanden Broek, C. J., Fairchild, R. C. and Elzinga, M. B.,
Improved ultrasonic nondestructive testing of pressure vessels, Nuclear
Regulatory Commission Report NUREG/CR-O135, May 1978.
 Real-time inspection of reactor ressels and piping components 41

7. Frederick, J. R. et al., Improved ultrasonic nondestructive testing of pressure

vessels, Nuclear Regulatory Commission Report NUREG/CR-0581,
January 1979.
8. Frederick, J. R. et al., Improved ultrasonic nondestructive testing of pressure
vessels, Nuclear Regulatory Commission Report NUREG/CR-0909,
September 1979.
9. Ganapathy, S. et al., Ultrasonic imaging techniques for real-time in-service
inspection of nuclear power reactors, Nuclear Regulatory Commission
Report NUREG/CR-2154, October 1981.
0. Ganapathy, S., Schmult, B., Wu, W. S., Hamano, N. and Bristor, D.,
Investigation of special purpose processors for real-time synthetic aperture
focusing techniques for nondestructive evaluation of nuclear reactor vessels
and piping components, Nuclear Regulatory Commission Report
NUREG/CR-2703, to be published.