DETECTION OF WEAK INTERFACE SIGNALS FOR SAME
MATERIAL BOND/WELD INSPECTION
B. A. Rinker1, E. E. Jamieson1, J. A. Samayoa1, T. G. Abeln2, T. P. Lerch3,
and S. P. Neal4
1
Honeywell Federal Manufacturing and Technologies, Kansas City,
MO 64141; Operated for the United States Department of Energy under
Contract No. DE-ACO4-01AL66850
2
Los Alamos National Laboratory, Los Alamos, NM 87545
3
Industrial and Engineering Technology, Central Michigan University,
Mount Pleasant MI 48859
4
Mechanical and Aerospace Engineering, University of MissouriColumbia, Columbia, MO 65211
ABSTRACT. In same material joining processes, remnants of the original interface may be acoustically
weak and difficult to detect ultrasonically in the presence of grain noise. For the nearly perfect (but
potentially inadequate) bond or weld the remaining interface may be defined only by small voids, cracks,
or by weak echo surfaces (grain boundaries) approximately aligned in the original interface plane. In this
study, we investigate the utility of correlation-based techniques for revealing the presence of interface
signals in the presence of grain noise as the signal-to-noise ratio is pushed below 1.0. These techniques
are attractive since they are independent of absolute scale and since the grain noise near the bond/weld
time window may be useful in establishing baseline results for a region without an interface (grain noise
only). Experimental results from two different same material joining processes are included in the study
along with supporting results from a simulation driven by experimental data.
INTRODUCTION
In the same material joining inspection problem, there is no acoustic impedance
difference between the two joined pieces. Previous studies have considered the detection
of cracks and voids in the original interface plane in the development of inspection
procedures [1-3]. In this study, we are most interested in the problem when there are no
such classical defects, and the only "defect" is a defect in the randomness of the scattering
sites. That is, the weld or bond may appear to be acoustically transparent at a signal-tonoise ratio (SNR) of less than 1.0, but the bond may be inadequate with grain boundaries
still approximately aligned with the original interface. This condition is sometimes called a
same material kissing bond. In the long term, the goal is to make ultrasonic measurements
and predict component performance whether the key issues are fracture and fatigue
resistance, corrosion resistance, or conductivity, for example. Our initial goals involve
detecting interface signals at SNRs of less than 1.0 and then establishing accept/reject
criterion based on the detection of these signals. This problem is a stochastic problem
where the NDE output will come, either implicitly or explicitly, as a likelihood that the
CP657, Review of Quantitative Nondestructive Evaluation Vol. 22, ed. by D. O. Thompson and D. E. Chimenti
© 2003 American Institute of Physics 0-7354-0117-9/03/$20.00
1080
calculated acoustic parameters are associated with grain noise - implying a satisfactory
weld with adequate recrystallization and grain growth across the original interface - versus
the likelihood that an interface signal is buried within the grain noise - implying a
potentially inadequate weld with a substantial fraction of the original interface still intact
Our approach will be to think of the problem not as individual rf signals which may
contain an interface signal, but as a set of rf signals which, if viewed as an image, may
show an edge or interface. We will key on the alignment of scattering sites near the
original interface. The acoustic impedance difference is still small (due to the same
material joining and the assumed absence of "classical defects"), but the approximate
alignment of scattering sites within a plane may lead to peak values in the "noisy signal" at
approximately the same time for some fraction of the ultrasonic measurement locations.
As a trial acoustic parameter to study interface signal detectability at low SNRs, we will
look at the correlation coefficient calculated between signals. In this paper, experimental
results are shown for inertial weld samples and upset-forge weld samples along with results
from a simulation study of the detectability of interface signals as a function of SNR.
EXPERIMENTAL PROCEDURES
Welded Samples
Experimental results will be shown from the ultrasonic inspection of 2 samples that
were upset-forge welded and 18 samples that were inertial welded. In both cases, the same
material joining process involved the welding of stainless steel to stainless steel. As
welded, all samples were of cylindrical geometry with approximate diameter and length of
0.75" and 1", respectively. In addition, whole, unwelded samples were used as grain noise
reference samples.
Ultrasonic Measurements
Ultrasonic measurements were made in an immersion mode using standard C-scan
type of raster scans with 1 mm spacing between measurement positions on a 31 x 31 grid
extending well beyond the sample. Each sample was scanned from both ends in separate
pulse-echo scans. A focused probe (1/2" diameter, 10 MHz) with nominal focal length of
4" was used with the water path set for focusing at the weld plane. Signals were digitized
using a 12-bit A/D card operating at 100 MS/s. As can be seen from the example A-scans
in Fig. 1, signals were digitized at each transducer location from prior to the front surface
reflection to beyond the back surface reflection. The left A-scan in Fig. 1 is representative
Grain noise
Interface signal
FIGURE 1. Typical measured signals in voltage vs. time A-scan displays.
1081
of a signal from a whole sample used to measure grain noise only. The right A-scan was
measured on a sample that was welded with welder settings chosen to produce an
inadequate weld, resulting in the large interface signal.
CALCULATIONS
An Experimental-Based Simulation
Using actual welded samples is clearly necessary, but creating large numbers of
welded samples, especially in the desired quality range where the weld is inadequate but
the SNR is still near or below 1.0, is very difficult and time consuming. This process is
compounded by difficulties in establishing destructively whether the weld is adequate or
not. In order to develop signal processing / detection approaches for this weak interface
detection problem, we have complimented the welded samples with a simulation. In the
simulation, measured grain noise signals (e.g., the grain noise in Fig. 1) and measured weld
interface signals (e.g., the interface signal in Fig. 1) are added in the computer to create
noise-corrupted interface signals as represented in Equation (1).
f simulated^
} = f grain noise (t,x>y) +B(^y)f interface^ ~
(1)
The SNR is controlled by scaling the interface signal prior to adding to the noise. The
scale factor, B(x,y), can be varied with lateral location to study the local detection
problem, and the interface signals can be given a systematic and/or random phase shift
(which can also be varied over xy) to simulate a non-planar interface. Example signals
from the simulation are shown for 3 different SNRs in Fig. 2.
Correlation Coefficient Calculations
The acoustic parameter considered in this paper is the correlation coefficient
calculated between signals. The correlation coefficient is attractive in that it is independent
of absolute scale and calculation of the correlation coefficient does not require
measurement system modeling and deconvolution to realize useful results. The correlation
coefficient does, however, show some dependence on the measurement system through
frequency content and beam field shape. In addition, grain size and morphology can
SNR = 0
SNR = 0.9
SNR = 2
FIGURE 2. Simulation output: example A-scans at 3 different signal-to-noise ratios.
1082
influence the correlation coefficient.
Correlation coefficient calculation were made on signals in a 7 x 7 grid centered
within the 3 1 x 3 1 C-scan measurement grid. This limited analysis set of 49 signals was
used in order to avoid edge effects. The correlation coefficient calculation was limited to a
time window surrounding the weld region. Each of the 49 signals was written into a 3D
array x(i9j9t) in MATLAB with the ij positions in the array corresponding to the xy
locations on the measurement grid and the 3rd dimension (k or time, f) containing the
vector of values representing each rf signal.
, j , t) - mx (i,
(2)
I [*(', J> *} ~ mx (', J)] J I [*(*', J + S,t)- mx (/, j + S)]
11*=*,
As shown in Equation (2), the correlation coefficient was calculated between signals in the
matrix with the spacing between signals controlled by the parameter S which was varied
from 1, for 1 mm spacing, up to 4, to yield 4 mm spacing between signals. This lateral
shift parameter is shown on the column index (/ + £) of the matrix, but was also used on
the row index (i + S) to allow calculation of the correlation coefficient for all possible
signal combinations at any S value. For a given scan, given S , and given SNR, an
average correlation coefficient, p , was calculated. For example, on a 7 x 7 measurement
grid with S = 1 , 84 values of p are calculated and averaged to yield p .
RESULTS
Simulation Results
For the simulation, 1 8 whole (unwelded) stainless steel samples were scanned on a
7 x 7 grid and used as grain noise scans. As discussed, interface signals from a scan 7 x 7
scan of a poor weld were scaled and added to these grain noise signals to create noise-
H
FIGURE 3. Simulation results: correlation coefficient vs. SNR (left graph) and correlation coefficient vs.
measurement position spacing (right graph) at various SNRs (bottom-to-top: 0, 0.3, 0.6, 0.9, 1.2, and 2.0).
1083
corrupted interface signals over a range of SNRs. This procedure yielded 18 sets of 49
signals at each SNR. Average correlation coefficient results are shown in Fig. 3. The left
figure shows the average correlation coefficient vs. SNR for each of the 18 samples at a
fixed signals separation of 1 mm (S =1). This figure shows that as the SNR increases, the
correlation coefficients for noise-corrupted interface signals separate from the correlation
coefficients for grain noise (SNR = 0). At SNRs of about 0.6 and above, the correlation
coefficient values are essentially separated from the correlation values for noise only. The
right graph in Fig. 3 shows the average correlation coefficient (circle data points) vs.
measurement position spacing, S, with each curve representing a different SNR. The
correlation coefficient for grain noise, which shows a mean at 0.3 for 1 mm spacing, drops
as we go to larger measurement position spacing. And, as expected, as the SNR increases,
the curve falls off more slowly - showing a residual correlation at increasing SNR and
increasing lateral spacing.
To take a first step into this problem as a stochastic problem, we used the mean and
standard deviation from the 18 observations at each SNR (see the left graph in Fig. 3) to
define a Gaussian distribution for the correlation coefficient for each SNR. At this point
we are using the Gaussian distribution only for visualization purposes. For the grain noise
only, the Gaussian curve is shown in solid line in both graphs in Fig. 4. The Gaussian
distributions for SNRs of 0.6 and 0.9, respectively, are shown in dashed line in the two
graphs in Fig. 4. In both cases, even at these SNRs which are below 1.0, the correlation
coefficient distributions for grain noise and for the noise-corrupted interface signal,
respectively, show a significant separation.
Experimental Results: Inertial Welds
Eighteen inertial weld samples were scanned from both ends, resulting in 36 scans.
At upper left in Fig. 5 is an example A-scan from one inertial welded sample. The SNR for
the interface signal for this example is around 1.0 with the average SNR across all scans at
about 1.6. The mean signals shown in lower left, calculated over the 49 signals in the 7 x 7
centralized grid, reveals the interface signal due to the non-random phase component of the
signal.
Given the nature of the mean signal, the average correlation coefficient at 1 mm
spacing should clearly show the presence of a signal which is relatively stationary in time
across the measurement grid. Instead of looking at grain noise measured from separate
unwelded samples to establish the grain noise correlation coefficient distribution, we use
each welded sample itself to look at the correlation coefficient for grain noise in time
FIGURE 4. Simulation results: correlation coefficient distributions.
1084
A-scan (taken near the axis of the sample)
Mean signal
15
20
Scan Number
FIGURE 5. Inertial weld results.
windows prior to and after the time window containing the weld. For each of the 36 scans,
we calculate 3 average correlation coefficients (for S = 1) - two for grain noise and one
for the weld window. The right graph in Fig. 5 shows the results with correlation
coefficients for the noise windows all plotting on the lower half of the graph. Clearly all of
the weld windows show a significantly higher average correlation coefficient than that for
any of the grain noise windows. Viewed on a scan-by-scan basis, the correlation
coefficient is lower in both grain noise windows than in the weld window. We could also
think of the grain noise windows as providing data to establish a grain noise correlation
coefficient distribution with the correlation coefficient for any weld window then compared
with this distribution.
Experimental Results: Upset-forge Welds
The 500x micrographs in Fig. 6 are for two samples welded using an upset-forge
welding process with two pieces of stainless steel being joined. The weld on the left is
basically a "perfect" weld with no apparent interface and a high level of recrystallization.
Shown below the micrograph is a typical rf signal (voltage vs. time signal) for a time
window that extends well beyond the weld region - even well beyond the region showing
recrystallization. In the center of the window, where an interface signal would appear, the
grain noise is relatively small, likely due to the smaller grain size in the recrystallization
region. The etchant used on the micrograph at right for a "marginal" weld, reveals the
presence of grain boundaries aligned with the original interface. In this case, there is very
little, if any, recrystallization. The etchant-enhanced interface viewed optically looks like it
must yield a large, easily detectable acoustic signal, but in fact, as can be seen in the rf
signal shown below the micrograph, there is no apparent acoustic interface signal. The
only "acoustic defect" in this welded sample is the alignment of grains boundaries with the
original interface, resulting in a SNR of less than 1.0 across the entire scan.
1085
FIGURE 6. Upset-forge weld micrographs and typical A-scans.
Calculation of the average correlation coefficient at 1 mm spacing for each sample
yields a "noise-like" correlation coefficient at about 0.25 for the "perfect" weld and an
uncomfortably high value of about 0.45 for the "marginal" weld. These values are
indicated along with the grain noise correlation coefficient distributions in Fig. 7. We
would expect the results of a hypothesis test to "detect" the interface signals in the
"marginal" weld by indicating a low probability that the 0.45 value is from a grain noise
correlation coefficient distribution.
Finally, we look at the average correlation coefficient vs. measurement position
separation for the 2 samples - using a grain noise window to establish the "grain noise
curve." These results are shown in Fig. 8. The curve for the "perfect" weld seems
FIGURE 7. Upset-forge weld results: correlation coefficient at 1 mm measurement position spacing.
1086
£
8 0.2
-0.2
-0.20
1
2
3
0
4
Measurement Position Spacing (mm)
1
2
3
4
Measurement Position Spacing (mm)
FIGURE 8. Upset-forge weld results: correlation coefficient vs. measurement position separation.
consistent with a noise only curve while the curve for the "marginal" weld falls off too
slowly and ultimately wanders about a residual value which is too far from zero to be
consistent with only grain noise.
SUMMARY AND DISCUSSION
Useful tools have been developed for studying the detection of weak interface
signals for same material weld inspection. Simulation results and experimental results on
welded samples are encouraging, even using a very simplistic correlation analysis
approach, with demonstrated ability to detect signals below a SNR of 1.0. As we move
forward, the scope of the simulation will be expanded and additional welded samples, with
quality near the accept/reject threshold, will be scanned. Ultimately, any approach will
need to be adapted for different geometries, materials, and inspection environments. We
also hope to move forward on critical, open questions associated with the engineering
materials version of tissue pathology - that is, based on a metallurgical analysis, is a given
welded sample or component acceptable or unacceptable?
ACKNOWLEDGMENTS
This research was performed with support under contract from Honeywell Federal
Manufacturing & Technologies, Kansas City, MO, and with support from the National
Science Foundation. A portion of this research was carried out while Brett Rinker was a
Research Assistant and Terry Lerch was a Postdoctoral Fellow in Mechanical and
Aerospace Engineering at the University of Missouri-Columbia.
REFERENCES
1. Nagy, P. B. and Adler, L., J. Nondestructive Eval 7, 3/4, 199 (1988).
2. Drinkwater, B., Dwyer-Joyce, R., and Cawley, P., in Review of Progress in QNDE,
Vol. 16B. eds. D.O. Thompson and D. E. Chimenti, Plenum, New York, 1997, p. 1229.
3. Taylor, J. O., in Review of Progress in QNDE, Vol. 16B. eds. D.O. Thompson and D.
E. Chimenti, Plenum, New York, 1997, p. 1215.
1087