Professional Documents
Culture Documents
Weld Inspection
...the consequences of decision-making can also have numbers attached to them, and
these two sets of numbers combined to solve the problem and determine the best decision
Lindley, 1971
6.1 Introduction
Inspection results from two or more NDT systems used for the examination of the same
specimen are usually different, sometimes in disagreement, even if related to the same
defect. In the case of two contradictory results, the NDT inspector has no reason to favour
one instrument over another and more inspections may be required using a third system to
confirm or refute the presence or size of a defect. In the worst cases, decisions could be
made resulting in unnecessary and expensive repair costs, or no action taken where a
major fault has been detected but wrongly characterised. Uncertainty and errors in signal
interpretation are factors which may cause problems in decision making. The information
from NDT sensors can be combined to help in decision making.
Both a theoretical approach to, and experimental results from, the application of NDT
data fusion to weld inspection are presented in this chapter. Non-destructive examination
of welds on plates, pipes and T-joints with induced defects performed using two
commercial eddy current systems is described. A parallel multisensor approach using
Bayesian statistical theory and Dempster-Shafer evidential reasoning was used to
combine defect information such as depth and length from up to five NDT sensors. Fusion
of eddy current, ultrasonic and radiographic data from a surface breaking defect is also
presented. In addition, NDT data fusion at pixel level is introduced. The performances of
these three data fusion processes applied to NDT are discussed.
The following procedure was adopted:
A 40.28 1.96
B 19.81 1.00
C 10.23 0.53
Root crack
Fig. 6.2 Cross-section of a butt welded plate with artificial flaws
6.3 Non-destructive examination of the test specimens
In addition to visual examination, dye penetrant inspection and MPI, weld samples have
been inspected using eddy current and ultrasound. The inspection results for these two
methods follow.
6.3.1 EDDYCURRENTTESTING
Two commercial eddy current systems, the Millstrong Lizard Topscan and the Hocking
Phasec 1.1, were used for inspection of the weld samples. For commercial reasons these
systems will not be referred to in the text by their respective names but as system A and
system B. Each sample was inspected and each defect detected was repeatedly sized
(depth, length and location were recorded) using each system in order to gather real data
for the fusion phase. It should be noted that in a real on-site inspection, only one numerical
datum from each instrument may be available and would be necessary for fusion. From the
27 defects present in our experiments, six defects were not detected by system B. In such
an event, one can see that no fusion of information is therefore possible as only system A
provided information. The defects undetected included root crack, an intermittent root
crack and a HAZ crack.
System A
Estimated Depth / mm
Actual Depth / mm
Fig. 6.3 Plot of the average estimated depth (and standard error) vs. actual depth for repeated
sizings of three calibration slots with eddy current system A
System B
Estimated Depth / mm
Actual Depth / mm
Fig. 6.4 Plot of the average estimated depth (and standard error) vs. actual depth for repeated
sizings of three calibration slots with eddy current system B
System A
Estimated Length / mm
Actual Length / mm
Fig. 6.5 Plot of the average estimated length (and standard error) vs. actual length for repeated
sizings of three calibration slots with eddy current system A
System B
Estimated Length / mm
Actual Length / mm
Fig. 6.6 Plot of the average estimated length (and standard error) vs. actual length for repeated
sizings of three calibration slots with eddy current system B
scattered and that the deeper the defect, the more scattered the data and the higher the
standard error associated with them. For both systems, the deeper the calibration slot, the
wider was the spread of estimated depth. The lowest error is 0.22 mm associated with the
depth of 0.57 mm (Fig. 6.7).
From analysis of estimated length against actual length of the calibration slots, it can be
said that system A provides very accurate measurements compared to system B. Small
standard errors associated with each value also show the high accuracy of system A over
system B. System B appears to overestimate defect length; this may be due to the design of
Actual Depth / mm
Estimated Depth / mm
Fig. 6.7 Diagram showing increasing error with increasing defect depth for measurements with
system A
the system, which does not provide the user with a reference point that is easily
identifiable. The automatic calculation, by the instrument, of the length and depth
measurements was not found to be sufficiently adequate and accurate. There is such a
disparity in the estimation of defect length between system A and system B that little
improvement is expected from the data fusion of length measurements. System A will
always have a higher weight than B associated with it and information from system B will
therefore be discarded. For this reason fusion of depth measurements will be investigated
in more detail than that of length. Moreover, in relation to fracture mechanics estimates of
structural integrity, the depth is an important parameter to measure accurately when
performing an inspection, as the inspector is required to calculate the degree of penetration
of the defect through the material and the remaining undamaged thickness, essential
factors affecting strength.
Defect
Defect Deoth / mm ^"B'*1 ' mm
Fig. 6.8 Normal probability density function for systems A and B for a 40.28 mm x 1.96 mm
calibration slot
transformed into the standard normal distribution by means of the equation
*-^ (6-2)
O
and the function O(z) can be plotted:
This transformation has the effect of centring the normal distribution at zero. The reason
for this transformation is that the probability associated with any defect depth can be
calculated using Z values and statistical tables. The graphs of normal probability density
function confirm the previous observations that as the defect depth decreases, system B
becomes more precise compared to system A, in terms of standard error and probability.
The relative probabilities associated with each system for different defect depths are
summarised in Table 6.2. It can be seen that the probabilities for each system increase,
although the probability for system B increases much faster than that for system A.
Therefore, more belief is given to the measurements of small shallow defects sized with
system B.
The x2 test
A x2 (chi-squared) test was performed in order to estimate the degree to which the signal
output from eddy current NDT systems A and B could be modelled using a Gaussian
normal distribution.3 The # 2 test was performed on the data collected from inspection of
calibration slots. The graphs resulting from the x2 test are shown in Fig. 6.9. It appears
from this test that the Gaussian normal distribution is a good approximation to model
sensor output and calculated probability associated with it.
The x2 test depends upon the characteristics of the sample distribution. It assesses the
difference between measured and expected values. The comparison between the measured
and expected values is made by calculating the x2 statistic, and is given by
I'-t^l <6.4,
/=1 Ci
where ra, and ef are the measured and expected values respectively.4 With large samples,
the x2 statistic approximates to a continuous x2 distribution. Comparison between the x2
statistic and a x2 distribution provides a measure of the probability of the distribution of
Fig. 6.9 Chi-squared test on calibration slots for depths of 1.96 and 1.00 mm for system A
P values
P values
Fig. 6.10 Normal plots of the expected values against the observed values for depth measurements
on calibration slots 1.96 (top), 1.00 (middle) and 0.53 mm long (bottom) for systems A and B
length measurements for system A does not appear to follow a normal distribution. This is
due to the lack of precision of the length measurements performed with system A leading
to a grouping of results due to rounding. In this particular case the ^ 2 test is more
appropriate to estimate the degree of normality of the set of sample values. However, there
is no evidence which shows that the distribution of length measurement does not fit a
normal distribution for the estimated length measurements from system B.
In general it can be said that the Kolmogorov-Smirnov test supports the possibility that
the spread of the measurements from NDT systems A and B does follow a normal
distribution.
Curve fitting
In order to apply the Bayesian and Dempster-Shafer theories to data from NDT systems A
and B, standard deviation values for any defect depth and length will have to be
Normal Plots - System A Normal Plots - System B
Expected Normal
Expected Normal
Expected Normal
Fig. 6.11 Normal plots of the expected values against the observed values for length measurements
on calibration slots 40.28 (top), 19.81 (middle) and 10.23 mm long (bottom) for systems A and B
determined. From the data collected during the inspection of the calibration slots, a
mathematical function which will model the expected standard deviations of both systems
was calculated. The curve fitting operation computed using a type function of the form
y = ax2 + bx+ c was selected. A standard deviation can be associated with any depth
measurement, and using Gaussian normal distribution tables, the associated probability of
this measurement can be calculated. This information is necessary in order to implement
the fusion phase which is described in section 6.4.
6.3.2 ULTRASONICEXAMINATION
Ultrasonic examination was performed on two ferritic samples using a Staveley Sonic 136
ultrasonic system and 45°, 60° and 70° probes. Three defects were studied: a slag inclusion
and a lack of side wall fusion (LOSWF) on a plate and a toe crack on a pipe. Each
Toe crack - P3075
Defect Length / mm
examination has been combined with eddy current and ultrasonic data using both the
Bayesian and the Dempster-Shafer approaches.