You are on page 1of 13

REVIEW OF DEFECT DISTRIBUTIONS FOR PROBABILISTIC

STRUCTURAL INTEGRITY ASSESSMENT

W. Zhao
Amey Vectra Ltd.
Warrington, UK

A. Stacey
Offshore Safety Division
Health & Safety Executive
London, UK

ABSTRACT A number of major NDE research projects have been


undertaken over the years with a view to determining POD
The defect distribution is a particularly important and POS distributions. Major NDE projects include PISC
parameter in the application of probabilistic defect [1-2], NORDTEST [3-10], NIL [11-14], work conducted by
assessment methods to the integrity assessment of the UCL NDE Centre for offshore structures [15], ICON
welded structures. However, the subject is complex and, [16-19] and TIP [20-21]. These projects were reviewed in
despite considerable research efforts in this area and the the SINTAP project [22].
availability of a significant body of data, this defect
distribution is subject to considerable variability and It is apparent that there is a need to collate information on
uncertainty. This paper reviews the data in the literature, available defect distributions and data with a view to
providing a status review of defect distributions and the identifying and developing suitable generic distributions
identification of areas requiring further investigation, with a for defect assessment. Some progress on the gathering
view to assisting with the future development of of information on defect distributions has been made in
probabilistic defect assessment methods. the SINTAP project. Further work is to be undertaken in
the current initiative to revise BS7910:1999 [23] and in the
FITNET project [24]. This paper reviews defect
INTRODUCTION distributions in the literature and makes a number of
recommendations for future work. Aspects covered
Defects are inherent to the welding process and can include initial defect size, initial aspect ratio, defect sizing
cause structural failure by fatigue crack growth and error, fabrication defects, probability of detection (POD)
fracture. Consequently offshore structures are subjected and surface / embedded defects.
to extensive inspection for weld defects during fabrication
and subsequently in service, including after repair. The
probability of defects being present depends on many INITIAL DEFECT SIZE DISTRIBUTIONS
parameters, such as fabrication techniques and quality
control, NDT methods and requirements, reliability of the
Probability distributions for initial microscopic
methods (i.e. probability of detection, POD, and sizing,
POS) and the operators. surface defect size

Probabilistic defect assessment methods are a useful tool Some key models of the initial microscopic surface defect
for the evaluation of structural integrity of welded size, based on exponential, Weibull and lognormal
structures. However, the effective application of defect functions, are listed in Table 1. Considerable uncertainties
assessment methods requires reliable information on the exist in these models. For example, the Marshal model
defect distribution and careful selection of the defect [25-26] gives results several magnitudes higher than that
distribution. Each of the factors determining the defect given by the Becher and Hansen [27] model in the
distribution in a component may be influenced by a probability of small defects.
number of basic variables. It is often not possible to
model all the variations. It is, however, desirable that any The Marshall Exponential model [25-26]
function used to describe the defect size distribution Among many models for the initial microscopic defect size
should be as accurate as possible at the tail of the distribution, the Marshall model is perhaps the most
distribution. typical one. It is expressed as an exponential form:

1 Copyright©2002 by ASME
1 a (1)
f a (a ) = exp( - )
am am
Distribution
where a is the defect size, Investigator Mean SD Occurrence
am is the mean defect size and equals 6.25 mm Type
(mm) (mm)
for the database considered in the Marshall Marshall Exponential 6.25 6.25 3.65 defect per
report. [25][26] vessel
Bokalrud and Exponential 0.11 0.11 16 cracks per
Karlsen [29] metre
A similar model is adopted in the R6 method [28]. Becher & Lognormal 1.81 1.78
Hansen [27]
Bokalrud and Karlsen [29] DNV [30] Weibull > 0.1
Bokalrud and Karlsen measured undercuts in butt-welded Recho[31] Lognormal 0.089 0.0887
plates of thickness from 10 to 25 mm and reproduced the SD – Standard deviation
depths of the undercuts on replicas from silicon rubber.
From the total of 827 randomly selected measurement Table 1: Initial microscopic surface defect size
points, 502 measurements revealed no undercuts. From distribution models
the remaining 325 locations an exponential distribution
with a mean value of 0.11 mm was fitted to the histogram Extreme distributions of the Initial defect size
of measured undercuts. The majority of the undercuts
were less than 0.5 mm deep. The use of extreme defect size distributions makes it
possible to focus on those larger cracks in a certain
The DNV Weibull model [30] population that are likely to cause failure.
A Weibull distribution is used:
The R6 model for the extreme defect size [28]
g -1 g The R6 method provides an extreme value distribution of
g æ a ö æ a ö (2)
f a (a ) = çç ÷÷ exp{ - çç ÷÷ } the defect depth derived from data obtained by extensive
am è am ø è am ø ultrasonic examination of Magnox reactor components. A
Gumbel Type II distribution is used:
am is set to be 0.1 mm. When g = 1, this becomes an
exponential distribution. æ æ a ö-B ö
f a ( a ) = A B Ba - B +1
exp ç - ç ÷ ÷ (5)
ç è Aø ÷
The Lognormal Model by Becher and Hansen [27] è ø
Becher and Hansen (1974) proposed a lognormal model
as
where a is crack depth,
1 é (ln (a/l ))
2
ù (3) A and B are constants.
f a (a ) = exp ê - ú
s aH 2p ë 2s 2 û
The DNV model [30]
In the DNV Weibull model as in equation 2, If small
with l = 1.3 mm and s = 0.82.
surface defects are believed to be present in any
significant number, then the prevailing crack will emerge
æ 1 æ t öö (4)
H = 1 - 0 . 5 e r fc çç ln ç ÷ ÷÷ from the largest or extreme defect. In this case, a more
è s 2 è l øø peaked initial distribution, with g > 1, is to be expected.

This gives a mean of 1.82 mm and a SD of 1.78 mm. Moan et al. [32]
More recently, Moan et al derived the mean initial defect
Recho [31] size and mean detectable defect size from defect data
Recho analysed NDT inspection results for 829 from more than 40000 in-service MPI and EC underwater
measurements of butt and fillet weld undercut depths and inspections as listed in Table 2. Moan et al compared
fitted to a lognormal distribution with an average value of their results with work by Bokalrud and Kalsen [29] and
0.089 mm and a standard deviation of 0.0887. Rocho found that the main difference comes from the sample
found that the “measurable” defects (a > 0.05 mm) only size. The data from Moan is based approximately on the
exist in 325 out of the 829 measurements. maximum defect size per three joints. Thus, it is an
extreme distribution per 3 joints by the transformation
through the following formula:

2 Copyright©2002 by ASME
E [ l 0 , max ] = l 0 {ln( 3 N j ) + 0 .557 } (6) study of Kountouris and Baker [37][40] and the study of
Rogerson and Wong [44] are obtained from the actual
where Nj is the number of defects per joint. offshore structure, the related detailed discussion is
presented in the section for fabrication defects.
For a joint with 1.4 metre weld, the number of cracks
would be 22.4, according to Bokalrud and Kalsen [29]. Bokalrud and Karlsen [29]
Moan et al [32] used the exponential distribution to model Bokalrud and Karlsen studied the data from UT test
results of a randomly selected weld length of 3200m out
the extremes, but they also argued that theoretically the
of 40,000 total weld length. A semi-normal distribution
Gumbel distribution should be used.
was fitted to the data. In addition, a semi-normal
distribution function for eccentricity is also proposed.
Distribution
Sample size
Type Mean (mm) Distribution
Exponential 0.94 Largest crack in three joints Investigator Mean SD Sample size
Type
Exponential 0.38 Largest hot spot crack per joint (mm) (mm)
Bokalrud & 3200 out of
Seminormal 0 2.88
Karlsen [29] 40,000 welds
Table 2: Extreme distribution models for Initial 1000m weld with
surface defect size Rogerson &
Weibull 1.81 1.78
defect rate of
Wong [44] 7 per 10 m
before NDT
Initial aspect ratio distribution
Kountouris &
Exponential 27,000 m
Baker [37] [40]
Very little information is available on the quantification of
the initial aspect ratio distribution of defects, particularly in Table 3: Initial embedded defect size distribution
offshore structures. It is important to make a distinction models
between this and the aspect ratio for in-service defects,
which is affected by the formation of larger cracks as a Defect Rates
result of fatigue crack coalescence. A few distributions
based on NDE test results obtained within the pressure Defect rates are important in the determination of the
vessel industry include: overall reliability of a structure. However, relatively little
(1) lognormal distribution of z =c/a by Harris [33] : attention is paid to its accuracy.

é æ æ 2
ù Harris [33-34] assumed that the number of cracks in a
öö
ê ç ln ç z z ÷ ÷ ú (7) body of volume V is a Poisson distribution with a mean
C è m ø ø
f a (a ) = exp ê - è ú
z 2p m ê 2m 2 ú
ê ú Pvm = 10-4 /in3 (9)
ë û
Therefore, the probability of having one or more defect is:
where zm = 1.336
C = 1.419 and m = 0.5382 Pv = 1- e-VPvm
(2) normal distribution of a/c by Bruckner et al [35] from (10)
NDE data with a mean of 0.5 and SD of 0.16.
The probability of n defects occurring in a weld under a
(3) Weibull model [28] of c/ a by the R6 method.
Poisson model is:
P ( X = n) = e - l ln / n! (11)
Lukjanov and Korobstov [36] however, proposed a
Weibull distribution model for defect length based on their
study of initial defect size and multiple coalescence: where l is the mean rate of occurrence of all sizes.

1 æ 2l + 2 ö
0 . 57
é æ 2l + 2 ö 1 . 57
ù (8) Marshall [26] found that there are 3.65 defects of all sizes
f a (2l ) = ç ÷ exp ê - ç ÷ ú per vessel.
3è 6 ø ëê è 6 ø ûú

Bokalrud and Karlsen [29] found the defect occurrence


Embedded Defects rate to be 16 defects per metre, including rather small
cracks. Rogerson and Wong [44] found that the defect
Much less information is available on probabilistic rate in a North Sea structure before final NDT and repair
distributions for embedded defects. Some major efforts was 0.7/m (i.e. 7 defects per 10 m of weld), reduced to
are summarised in Table 3. As the defect data for the 0.2/m after final NDT and repair.

3 Copyright©2002 by ASME
Defect Weibull Parameters NDE
Joint Investigator
Type Range a b e Method
Planar 5.1 1.4 0.1
Embedded as
Non-planar 0.93 0.7 0.1
fabricated Tubular joints MPI and Rogerson &
All 3.3 0.95 0.1 1 km welds UT Wong [44]
Embedded After NDE and
1.22 0.8 0.1
repair
Surface breaking crack depth 1.85 0.85 0.1 Ship plate Becher &
X-ray
Slag inclusion crack length 7.5 0.45 2.0 Hansen [27]
Depth 7 2.18 0
Lack of fusion Tubular joints
Length 3.41 2.14 0 UT Burdekin &
Depth for 1.9
Lack of Depth 8.13 2.76 0 Townend [43]
km welds
penetration Length 3.61 3.65 0

Table 4: Parameters of the Weibull model (Equation 12) for fabrication defect distributions

Kountouris and Baker [40] found that the defect rate, as a Another set of data was obtained from the results of a
ratio of defect length to the total length of weld, lies in the separate NDT inspection on one of the four structures,
range 0.010 to 0.014. They also found that defect rates for there was a wide scatter in the number of defects per
different yards are found to be radically different. The metre. The data were from ultrasonic testing of 250 m of
amount of NDE was found not to affect the defect rates. welding. Exponential and Weibull distributions were fitted
to the defect length. However, the Weibull distribution
was considered to be a better fit to the data than the
FABRICATION DEFECTS exponential one as (a) it described all the data by the
same function type and (b) gives higher estimates than
Records from fabrication sites often show different the exponential distribution for the large defects and is
distributions, as different factors are present. For therefore more conservative.
instance, some results incorporate NDT reliability and
even defect repair. Nevertheless, in general the Weibull Dufresne [42]
distribution shows a good fit with data with the lognormal Dufresne analysed NDE results (UT and X-ray) obtained
distribution as the second favourite. While the initial from three European nuclear pressure vessel
defect distribution discussed above provides models for manufacturers before repair. The data on defect length
application at the microscopic level, the fabrication defect show that the number of defects per weld varies between
results provide useful support information for uncertainty 0-50 and that the lognormal distribution is a good
modelling in specific or macroscopic applications. It has approximation for defects larger than 20 mm. The
been found that a general Weibull distribution in the form distribution for defect depth was not specified.
of equation 12 will fit the data well. Some of the Weibull
parameters are listed in Table 4. Burdekin and Townend [43]
Burdekin and Townend analysed ultrasonic NDT records
æ a -e b ö to determine the distribution of defect heights (lack of
Fa (a) = 1 - expç - ( ) ÷ (12) fusion and lack of penetration). The records were for
è a -e ø 100% ultrasonic NDT of 1.9 km of 440 manual metal arc
T-butt welds at tubular node intersections in a series of
Becher and Hansen [27] jacket structures fabricated from BS 4360 grade 50D
Becher and Hansen (1974) identified 242 defects from steel. For these data, a Weibull distribution was found to
347 radiographs (X-ray) corresponding to 72 m of welding give a moderately good fit with a cumulative probability of
performed on hull plate with 22 to 26 mm thickness. They values equal to or less than size a (mm) given by
fitted Weibull distributions to the length and depth
dimensions.
é æ a ö 2 . 39 ù (13)
F a (a ) = 1 - exp ê - ç ÷ ú
Rodrigues and Wong [41] ëê è 7 . 38 ø úû
Rodrigues and Wong presented an analysis of data
obtained from node welds in four different North Sea
structures fabricated by different European fabricators.

4 Copyright©2002 by ASME
Rogerson and Wong [44] reportable and repairable defects and further into pre-
Rogerson and Wong carried out a detailed analysis of repair and post repair categories. It was also separated
embedded weld defects namely defect height, for one according to inspection methods and whether the defects
particular structure in the North Sea and the survey were planar or non-planar. Among the findings, the
included about 1000 m of welding for 12 vertical node following two are particularly worth noting:
joints and 6 horizontal node joints in the important splash
· the probability distributions for defect size were
zone region of the structure. The data were obtained by
generally similar to those reported in the literature.
magnetic particle inspection and ultrasonic inspection
The Weibull distribution gave a better fit but
methods and were divided into planar and non-planar
sometimes there was little difference between the
defects.
different distributions.
Notably, Rogerson and Wong combined the study of · the effect of allowing for POD was to increase
defect distribution with failure probability determination in significantly the estimated number of smaller defects
the context of NDT and repair. The Weibull distribution present in the jacket with an overall increase in the
gave a good fit for both planar and non-planar defect numbers of defects and a reduction in the mean size
height, particularly at the extreme values, deemed to be and standard deviation of defects.
the most important in their study. They concluded that
the NDT procedure should be only required to have a high
and consistent probability of detection of large defects. PROBABILITY OF DETECTION (POD) DISTRIBUTIONS
This, however, should be viewed in the context of the
technological limit and requirement in both POD and Inspection reliability is an important assumption. Usually,
reliability analysis at the time. tests conducted in the laboratory to determine the
probability of defect detection (POD) can only model a
Baker and Kountouris [37][38] limited range of factors:
Baker and Kountouris analysed surface breaking defects · inspection technology
found by magnetic particle inspection in the lower hull of · environmental condition, with the following three
the Conoco Tension Leg Platform. The length of the weld levels:
inspected was of the order of 50 km. Weibull and 1. laboratory
lognormal distributions were fitted to the data with an 2. in air
explanation of the variations in defect sizes under the 3. underwater
influence of defect location, defect type, weld type, · existence of multiple cracking or defect. Everything
positions in the structure, fabrication and welding process. else being equal, a part with two cracks is more likely
Analysis was undertaken on defect length, depth and to fail the inspection than a part with only one crack.
aspect ratio. · human factors
· mathematical uncertainties from the resultant crack
Baker and Kountouris also analysed embedded defects
PDF after detection or inspection
found by ultrasonic testing for a weld length of the order of
· other geometric factors, e.g. plate thickness, crack
27 km. The results indicated that the Weibull distribution
orientation, etc..
gave on average a better fit than the lognormal
distribution.
Mathematical POD models
Baker [37] indicated that the occurrences of defects from
fabrication are dependent primarily on: Given the complex nature of the detection probability,
there are few numerical variables for the quantification of
· weld type (joint geometry). uncertainty. Often, only the defect size is the only known
· weld process e.g. SAW, SMAW and GMAW. numerical value. Some models contain variables to
· positions in the welds. quantify the inspection quality with limited success.
· welding procedures Let POD(a) be the detection probability for crack larger
than the crack size a. Most POD models are in the form
Thurlbeck [45]
Thurlbeck analysed fabrication defects obtained from the
POD(a) = 0 for a < ath,
fabrication inspection records of two offshore steel jackets
in the North Sea fabricated in 1986 and 1991. The data POD(a) = f(a) for ath < a < au, (14)
consisted of the results of magnetic particle inspection,
POD(a) =1 for a > au,
ultrasonic and radiographic testing of some 3800 welds of
total length of about 9.35 km. The data were divided into

5 Copyright©2002 by ASME
where ath is threshold of detectability and for the PVRC data the least square fit was
au is the size with 100% detectability
POD (a ) = 1 - exp{ - (a / 6 . 3 )} (19)
If ad denotes the smallest detectable crack, the probability Townend [49] also stated that to account for the effects of
of detecting a crack of size a is equal to the probability the defect detection function in the estimation of the
that ad is smaller than a, i.e. defects in the structure prior to NDT and repair, a discrete
approximation can be applied. He demonstrated this by
F a d ( a ) = POD ( a ) (15) the following example, if 20 defects have been detected in
the 10 mm to 20 mm size range and the detection
The POD curve is therefore identical to the distribution of function gives a mean probability of detection for this size
the smallest detectable crack size. In study carried out by of 0.8 then the expected actual number in this range
Marshall [25], the function name, B(a), is also used for the would be 25 defects.
probability of non-detection.
DNV curve [30]
The Exponential Model DNV introduced a series of POD curves as follows:
Marshall [25] gives the probability of not detecting a as b
P(2c) = 1-1/(1+(2c/ X0) ) (20)
B(a) = 0.005 +5.53 exp(-0.16 * (a- ath)) (16)
where X0 and b depend on the inspection scenario with
ath is assumed to be zero. typical values listed in Table 6:

Madsen[47] used a shifted exponential distribution to Inspection scenario X0 b


model the POD as follows: MPI under water 2.950 0.905
MPI above water; ground test surface 4.030 1.297
æ a - a min ö MPI above water; not ground test 8.325 0.785
POD ( a ) = 1 - exp ç - ÷ (17)
l surface
è ø
for a > a min. . Eddy current 12.28 1.790
Table 6: DNV POD parameters
The parameter l is a measure of the quality of inspection
and has a value between 0 (perfect) and infinite (poor). Harris Lognormal model [33-34]
Some of the empirical factors of l are shown in Table 5. Harris used the data reported by Tang [50] for not
detecting a crack of size a. The data reported was in the
UCL has also been quoted to use l = 1.2. (Kirkemo [46]).
range a < 2.5 mm. Harris compared Weibull, lognormal
and Gamma function distributions and proposed a
Location lognormal model for crack POD in reactor piping:
Investigator Test condition Tubular in Butt weld
sea water in air é
1 æ a öù (21)
Lab TSC ACFM 0.7 0.21 POD ( a ) = 1 - erfc ên ln ç * ÷ ú
2 ë è a øû
Lab MPI 0.89
UCL [15] Lab Comex/Hocking
2.08 0.32
Eddy Current where a* is the crack with 50% chance being detected
Kirkemo [46] 1.0 – 1.4
Madsen [47] underwater MPI 1.3
Harris suggested a* = 1.6mm and n = 1.33.
MPI and EC at in-
Moan [48] service and sea 1.95 1.- 1.2
condition Harris [34] pointed out that the probability of not detecting
a surface defect would decrease as the defect size
Table 5: Quality parameter l for the POD curve increases and is nearly unity for small defects whereas it
asymptotically approaches zero with increasing defect
Townend [49] modelled and compared the performance of size. However, it is never equal to zero as there cannot
the PISC and the PVRC programms by minimizing the be absolute certainty that a defect will be detected. A
least squares errors. The best fit for the PISC data was modified expression, which also accounts for the two-
found to be dimensional nature of defects, is presented below:

POD (a ) = 1 - exp{ - (a / 57 )} (18) B (A ) = e +


1
2
(
(1 - e )erfc n ln A A * ) (22)

6 Copyright©2002 by ASME
Statistical comparison of POD models
A* is the crack area which has 50% detection probability,
e is the probability of non-detection of very deep cracks Wolstenholme [54] compared different models for
and n is a parameter which controls the slope of the probability of detection curves for the mathematical
POD(a) line. uncertainties. Three distribution types are compared: 1)
exponential, 2) Weibull, and 3) log-odds. It is found that
Taking into consideration the false call probability, the exponential model remains adequate for the modelling
Heasler, Taylor and Doctor [51] proposed a logistic model of POD but it is less versatile than other multi-parameter
as follows: models. Wolstenholme also suggested that POD models
should be able to include false reporting.
-1
POD(a) = (1+ exp (-(a + b a))) (23)
Human Factors correction
Thus,
Human factors are more complex to define. Apart from
FCP = (1 - exp(-a))
-1
(24) competence of the inspector, there are also the elements
of expectation and correlation between different
Berens and Hovey [52] suggested a log-odds model for inspectors. Wall et al [55] applied a human reliability
the probability of leaving a defect in a vessel of the form correction to the theoretical POD as:

PODactual = POD model * H (28)


B (a ) = 1 (25)
(1 + exp( a + b ln a )
where H is the human error of POD.
where a is the defect size and a and b are constants.
A general equation was also suggested in a European
Multiplication models American workshop [56] as:
Dufresne [42] stated that the longer the defect, the greater
the number of crossings of the UT transducer and hence PODactual = PODintrinsic – g(application technique) –
the risk of missing a defect is reduced, such that - h(Human factor) (29)

Pn=1-(1-P1)2b/f (26)
DEFECT SIZING ERROR MODELS
where P1 is the probability to detect at each crossing
Pn is the probability to detect at n crossing The measurement of defect dimensions by a non-
2b is the defect length destructive examination device requires skill and is
f is the diameter of the beam. governed by a large number of uncertainties and
limitations. The accuracy by which such measurements
The beam diameter is taken as 40 mm, and the 9 mm are made is valid only for particular categories of defect
defect length corresponds to 50% detection. but information about the tendency to undersize or
oversize defects can be obtained by examining trends
Temple [53] gave a model that includes many variables, between measured and actual sizes or means and
several inspection techniques Q and human variability. It standard deviations of sizing errors (measured size -
is given by actual size), when data is available.

é ù Sizing errors can have serious consequences since, if


1 + exp ê - å a * ú significant defects are overestimated, this could result in
êë i i ú
BQ = û (27) unnecessary repairs causing financial losses and possibly
[ (
1 + exp å i a i x - x * i )] lead to the introduction of further defects by the repair
process. Conversely, if significant defects are overlooked
this could have serious effects on overall structural
where there are i parameters in x. The parameter integrity. From a reliability analysis for the pressure
ai determines the sharpness of the boundary between vessel with flaw, Simonen and Khaleel [57] have shown
acceptance and rejection, the parameters xi* are the that, although defect detectability is the most limiting
values at which the probability of correct rejection reaches factor to reduce failure probability, the gross error in sizing
0.5 if there was only that parameter present. can negate the expected benefit of inspection methods.

7 Copyright©2002 by ASME
Jessop [58] used UT for flaw sizing in welds and found Errors about this line were normally distributed with a
that COV is about 20% and 30%. The worst case method mean of 7.2 mm and a standard deviation of 6.28 mm. It
(a standard dB drop) exhibited the poorest accuracy with was found that small defects were overestimated in size
a COV of 0.3. Other analyses of similar UT whereas large defects were underestimated in size. The
measurements from less carefully controlled testing crossover from overestimating to underestimating
situations produced COVs of double the above value. occurred at 25.67mm.
COVs of 0.2 and 0.3 are, however, appropriate for
validated automatic or closely prescribed manual Haines et al [59] confirmed the finding by Townend in the
ultrasonic testing procedures when using suitably qualified review for PISC II round robin test but he also pointed out
and experienced personnel (as found in the nuclear power that undersizing occurs more frequently than oversizing
industry). for rejectable defects.

Defect size errors can be modelled by either of the Kountouris and Baker [39] analysed data on fatigue
following terms: cracks from four typical tubular welded joints published by
the UK Department of Energy. All cracked specimens
am = b1* ab2 * e (30) were detected by magnetic particle inspection and
or subsequently sized using manual ultrasonic and other
am = b1+ b2. a + e (31) techniques. The true dimensions of the defects were
evaluated by destructive sectioning. It was found that for
with e representing errors whilst b1 and b2 are obtained both defect depth and length there is one measurement
from regression analysis. that deviates significantly from the rest of the data. A
linear regression analysis of the measured dimensions
Heasler et al [51] found from PISC-II data that as and the true dimension was performed on all four cases
standard deviation does not go to zero when the defect (including and excluding the extreme observations).
size is approaching zero, hence the second
representation of am is more appropriate. It was found from the results that, for the case of defect
depth in the two cases (including and excluding the
Based on surface defects in a weld detected by an extreme observation), sizing errors are dependent on the
ultrasonic device on aluminium specimens, Tang [50] actual defect size. Defects of depth up to 7 mm were
suggested the following calibration curve between actual found to be overestimated whereas defects greater than
and measured defect depth. this value were underestimated. For the defect length,
the reliable case was found to be the one that excludes
E (C a | C m ) = 0 . 0592 + 0 . 839 C m for 0 < C m £ 0 . 2 in (32) the extreme observation and in this case the defects were
usually overestimated.
This is a normal random variable with zero mean and Osage, Wirsching and Mansour [60] used data from crack
0.0342 inches standard deviation. The specimens were tip diffraction measurements of defect depth to model the
precracked in fatigue to produce a small surface defect depth using a lognormal distribution. The mean
thumbnail crack, they were inspected by NDT to defect sizes were assumed to be located at the midpoint
determine the sizes and locations of the flaws, then tested between the upper and lower limits of the reported
to failure in a uniaxial tension test. The failed surface was measurements. The standard deviation was set equal to
examined to determine the actual size of flaws. the upper limit minus the mean value divided by 3. The
resulting COV was found to be a function of mean defect
Townend [49] looked at data of lengths of rail containing depth:
defects that had been sized ultrasonically before being
removed from track. The actual defects were made
am = (Upper Limit + Lower Limit)/2.0
visible by breaking the length of rail open on a large
SD(a) = (Upper Limit - Lower Limit)/6.0 (34)
hydraulic press and then the actual defects were
measured using a steel rule. The data was then tested
The COV was found to lie in the range 0.2 - 0.3.
for correlation between ultrasonic sizing error and defect
height in mm and linear regression analysis of the data
gave a least square line of sizing error versus actual
EQUIVALENT INITIAL DEFECT SIZE
defect size as,
The information revealed large diversity of the initial
Error = 13.86 – 0.54 (actual size) (33)
defect size distributions from various investigators. Very
often, the fundamental defect size distribution cannot be

8 Copyright©2002 by ASME
verified by normal NDT methods. Recho [31] proposed a (1) The initial microscopic defect distribution tends to be
method to backtrack the initial defect size distribution from exponential but there is considerable uncertainty in
fatigue experiment results of welded joints by using LEFM the distribution parameters.
and Miner’s sum.
(2) The defect distributions from fabrication records are
more likely to be Weibull distribution. The difference
Palmerg et al [61] used the concept of equivalent initial
to initial microscopic defect distribution can be due to
flaw size in the assessment of aircraft structures, based
on a reversal of the time to crack initiation (TTCI) (a) the inspection efficiency on smaller defects
distribution, which itself is assumed to be a Weibull (b) sampling effects
distribution. Palmerg et al.[61], however, identified three (c) lack of growth in small defects by fatigue
problems when using this method: (1) applicability for the threshold
short crack range, (2) the threshold value, and (3) this
model may be different for different design variables and (3) The distribution of extreme defect size can be
load spectra. Therefore, it is not an inherent material modelled by the Gumbel distribution.
representation.
(4) The defect rate is subject to great uncertainty.
Yazdani and Albrecht [62] used the AASHTO allowable S- Furthermore, different researchers used different
N curve tests data and integrated the Paris crack growth definitions of defect rates, e.g.:
equation backwards to obtain the equivalent initial defect -4 3
size and subsequently fitted a lognormal distribution to the · per volume: 10 / in - Harris [34]
data. · per pressure vessel: 3.65 - Marshall [26]
· per metre: 16 - Bokalrud and Karlsen [29]
This equivalent initial defect size approach has the benefit
of relating to the final damage status and using the · length per metre of weld: 0.01 to 0.014 – Baker
existing information in experiments without knowing the [38]
precise defect distribution. However, apart from the · per joint: 1/3 - Moan [32]
problems raised by Palmerg, the S-N curve and the
fracture mechanics parameters are also subjects of great (5) In general, the POD curve follows a monotonic
uncertainties. incremental distribution function and is affected by
many variables. The simplistic exponential function
still represents an adequate form. However due
DISCUSSION considerations should be given to inspection quality,
human errors and multiple probabilistic events.
An extensive review has been carried out of defect
distributions for reliability analysis. The sketch in Figure 1 (6) Defect sizing errors can adequately be represented
shows three generalized distribution models for defect by either normal or lognormal distributions, being
size. From the review in this paper, the following points functions of defect size and category and inspection
can be made: methods. Small defects tend to be oversized while
large defects tend to be undersized. The COV of
Initial defect distribution fa(a0) sizing error can be up to 0.3.
modelled by Exponential
Probability density function

(7) Given the uncertain nature of the defect distribution,


an equivalent defect size distribution obtained by
Detected defect distribution fa(a0) x backtracking experimental results may give some
POD(a) modelled by Weibull practical results. The disadvantage is that the basic
parameters in LEFM and S-N curves methods used
Extreme defect distribution in the backtracking are themselves subjects of
modelled by Gumbel considerable uncertainty.

Defect size a
Figure 1: Illustration of defect size distribution
models

9 Copyright©2002 by ASME
FUTURE STUDIES Model uncertainty of POD curves
Given the simplistic form of the POD curve, the POD
It is apparent from the work reviewed above that further curve itself has a large model uncertainty. The principal
work is required to develop defect distributions that can be challenges are:
used with greater confidence. Areas considered to merit Ÿ identification and quantification of model uncertainties
further attention are described as follows.
Ÿ reduction of model uncertainties, e.g. by database to
match the POD curve to on-site conditions.
Sensitivity study of defect size distributions
Defect size population in welds can be divided into
approximately three types of defects, small, medium and
CONCLUSIONS
large:
Ÿ small defects make a large percentage of the initial Probabilistic structural integrity assessments are highly
defect size population. Under fatigue loading, these sensitive to the defect distribution data and consequently
defects will either coalesce to form larger cracks or a significant amount of research on the subject of NDE
stop growing below the threshold. distributions has been, and continues to be, carried out.
Ÿ medium size defects can remain undetected and grow The extent of the research reflects the complexity and the
to form large cracks following the Paris Law. uncertainties associated with the subject.
Ÿ large defects are most likely to be detected and
rejected from initial inspections. Hence, their initial This paper has presented a review of available
distribution may be truncated. However, the information on defect distributions and made
consequence of missing the defect will also be recommendations for future work with the aim of
relatively large. facilitating the development of guidance for incorporation
into defect assessment codes.
This review has shown different defect size distributions
from fabrication records and from initial defect
measurements primarily due to uncertainties in small ACKNOWLEDGEMENTS
defect. However, failure probability may be dominated by
larger cracks. The authors wish to thank the Health and Safety
Executive (HSE) for permission to publish this paper. The
In order to derive generic form of defect size distributions views expressed are those of the authors and not
for practical application, the sensitivity of structural necessarily of the HSE.
reliability to the uncertainties of distribution parameters
especially in small defect size region should be
understood. The sensitivity study should be carried out REFERENCES
with due consideration to all relevant variables, e.g. defect
rates, fatigue growth, size effects of fracture toughness [1] PISC-: Nichols, R.W. and Crutzen, S., Ultrasonic
etc. inspection of heavy section steel components, The
PISC II final report, Elsevier Applied Science,
Harmonisation of description of defect size distributions Barking UK, ISBN 1-85166-155-7, 1988.
This review has shown that although a large amount of [2] PISC-: Lessons learned from PISC-III, Report No
research has been carried out on defect size distribution, EVR 16366 EN, Draft, 1/2/96. PISC-I: Evaluation of
some of the data are of limited use as they have been the sizing results of I'2 flaws of the full scale vessel
obtained under different assumptions. Figure 1 shows installation, PISC III report No 26 - Action 2 - Phase I,
three different types of distribution for defect size with JRC report No EUR 15371 EN, 1993
three different sampling methods. It is therefore
necessary to develop uniform procedures for the [3] NORDTEST: Førli, O., Development and
determination of defect size distributions. optimisation of VDT for practical use – Nordtest NDT
programme - project presentation, 5e iVordiska NDT
Investigation of defect rates Symposiet Esbo, Finland, IIW Report Number 1IW-
The overall failure probability of a structure is the product V-967-91, 1990.
of defect rate and the conditional failure probability for [4] NORDTEST: Førli, O., Development and
given defect distribution. However, the quantification of optimisation of NDT for practical use – Optimal NDT
defect rate is far from certain in current literature. efforts and use of NDT results, 5e Nordiska NDT
Symposiet Esbo, Finland, IIW Report Number IIW-V-
968-91, 1990

10 Copyright©2002 by ASME
[5] NORDTEST: Førli, O., Development and [19] ICON: Rudlin, J.R. and Dover, W.D., Performance
optimisation of VDT for practical use – Nordtest NDT trends for POD as measured in the ICON project,
programme - project presentation, 5e Nordiska NDT Proc. 1996 OMAE, Vol. II, p.509-513, 1996.
Symposiet Esbo, Finland, IIW Report Number 1IW-
[20] TIP: Rudlin, J. and Austin, J., Topside inspection
V-967-91, 1990.
project: Phase I Final report; Offshore Technology
[6] NORDTEST: Førli, O., Development and Report OTN 96 169 Nov. 1996
optimisation of NDT for practical use – Optimal NDT
[21] TIP: Rudlin, J., Myers, P. and Etube, L., Topside
efforts and use of NDT results, 5e Nordiska NDT
inspection project: Phase II Final report; Offshore
Symposiet Esbo, Finland, IIW Report Number IIW-V-
Technology Report OTN 96 169 Nov. 1996
968-91, 1990
[22] SINTAP Structural Integrity Assessment Procedure
[7] NORDTEST: Førli, O., Development and
for European Industry, BRITE EURAM BE 95-1426,
optimisation of NDT for practical use - Reliability of
Contribution to TASK 3.4: Compilation of NDE
radiography and ultrasonic testing, 5e Nordiska NDT
effective data”, SINTAP Task 3.4, Crutzen, S. et al,
Symposiet Esbo, Finland, IIW Report Number IIW-V-
1999
969-91, 1990
[23] BSI , BS 7910 “Guide on methods for assessing the
[8] NORDTEST: Kauppinen, P. and Sillanpää, J.,
acceptability of flaws in metallic structures”, 1999
Reliability of magnetic particle and liquid penetrant
inspection, IIW Report Number IIW-V-970-91, 1990 [24] FITNET, A European Fitness-for-Service Thematic
Network, sponsored by the European Union.
[9] NORDTEST: Kauppinen, P. and Sillanpää, J.,
Reliability of surface inspection techniques for [25] Marshall W., "An Assessment of the Integrity of PWR
pressurised components, SMIRT 11 Transactions Pressure Vessels," Report by a study group,
Vol.G No GI5/5, Tokyo, August 1991. available from H.M. Stationary Office, London, United
Kingdom, 1976
[10] NORDTEST: Kauppinen, P. and Sillanpää, J.,
Reliability of surface inspection techniques, Proc. [26] Marshall for UKAEA, “An assessment of the integrity
12th World Conf. on Non-Destructive Testing, of PWR pressure vessel”, Second report by a study
Elsevier Publ. Amsterdam, 1989SMIRT 11 group under the chairmanship of Dr. W Marshall,
Transactions Vol.G No G15/5, Tokyo. August 1991 March, 1982
[11] NIL, Evaluation of some non-destructive examination [27] Becher, P. E. and Hansen, B., “Statistical evaluation
methods for welded connections with defects, NIL of defects in welds”, danish Institurte of Welding,
report NDO 86-23, 1986 (in Dutch) 1974
[12] NIL, Advanced flaw size measurement in practice, [28] Harrison R. P., Loosemore, K., and Milne I.,
NIL report GF 91-04, 1991 (in Dutch) "Assessment of the integrity of structures containing
defects." CEGB Report R/H/R6, Central Electricity
[13] NIL, Non destructive testing of thin plates, NIL report
Generating Board, U.K., 1976
NDP 93-40, 1995. NIL, NDT of thin plates -
evaluation of results, NIL report NDP 93-38 Rev. 1, [29] Bokalrud, T. and Karlsen, A., “Probabilistic fracture
1995 (in Dutch) mechanics evaluation of fatigue failure from weld
defects in butt weld joints”, Proc. Conf. on fitness for
[14] NIL, NDT-Regulations, NIL Report NDP 95-85, 1995
purpose validation of welded constructions, London,
[15] UCL: Visser, W., Dover, W.L. & Rudlin, J.R., Review paper 28, 1981
of UCL underwater inspection trials, HSE OTN 96
[30] DNV, DNV Classification Notes No. 30.6, “Structural
179, 1996
Reliability Analysis of Marine Structures”, July 1992
[16] ICON: Project "ICON", Final Report, Contract No
[31] Recho, N., “Potential use of fracture mechanics in
OG/00098/90/FR/UK/IT, EC DG XII Programme
the fatigue design of fillet welded joints”, Proc. 5th
THERMIE, Report No S.94.006.03, Issued by
European Conf. on Fracture, Lisbon, Vol. 1, 1984
IFREMER, 12/94.
[32] Moan, T., Zhong W., and Vardal, O. T., “Initial crack
[17] ICON: Offshore Technology Report OTN-96-150,
depth and POD data based on underwater inspection
Intercalibration of offshore NDT (ICON), Commercial
of fixed steel platforms”, IOCOSSAR2001, Newport
in confidence PENIS/2736, HSE, August 1996.
Beach, June 2001
[18] ICON: W.J. and Rudlin, J.R., Defect characterisation
and classification for the ICON inspection reliability
trials, Proc. 1996 OiVAE, Vol. II, p.503-508, 1996.

11 Copyright©2002 by ASME
[33] Harris D.O., "Probabilistic Fracture Mechanics," offshore steel jacket structures”, Report for the HSE,
Pressure Vessel and Piping Tech, Ed. by C. 30 April 1996
Sundararajan, 1985, pp. 771-791.
[46] Kirkemo F, “Applications of probabilistic fracture
[34] Harris D.O and Lim E Y, "Application of a mechanics to offshore structures”, Appl Mech Rev
Probabilistic Fracture Mechanics Model to the Vol. 41, No. 2 ,1988
Influence of In-service Inspection on Structural
[47] Madsen H.O, Sorensen J D, “Probability-based
Reliability", ASTM STP 798, 1983
Optimisation of Fatigue Design, Inspection and
[35] Bruckner A., Haberer R., Munz D., and Wellein R., Maintenance”, Integrity of Offshore Structure,
"Reliability of the Steel Containment of a Nuclear Glasgow, 1990
Power Plant Using Probabilistic Fracture
[48] Moan, T., “Reliability and risk analysis for design and
Mechanics," Paper presented at ASME Pressure
operations planning of offshore structures, Structural
Vessel and Piping Conference, Portland, Oregon,
safety and reliability”, Schueller, Shinozuka & Yao,
June, 1983.
ed., ISBN 90 54 10 357 4, 1994
[36] Lukjanov, V. F. and Korobstov, A. S., “Imitative
[49] Townend, P. H., “Statistical aspects of fracture of
simulation of welded joint fracture under low-cycle
weld defects in the node joints of tubular steel
loading”, Int. Press. Vess. Piping, Vol. 47, No. 2,
offshore structure”, PhD thesis, UMIST, Oct. 1981
1991, pp 193-206
[50] Tang, W. H., “Probabilistic updating of flaw
[37] Baker M. J., Kountouris I. S. and Ohmart R. D.,
information”, J. of Testing and Evaluation, JTEVA,
“Weld Defects in An Offshore Structure, A Detailed
Vol. 1. No. 6., pp 459-467, 1973
Study”, 1988 Int. Conf. On Behavior of Offshore
Structures, Trodheim, June 1988 [51] Heasler, P. G., Taylor, T. T. and Doctor, S. R.,
“Statistically based reevaluation of the PISC-II round
[38] Kountouris I. S. and Baker M. J., “Defect
robin test data”, NUREG/CR-4908, Washington D. C.
assessment: analysis of the dimensions of defects
1993
detected by magnetic particle inspection in an
offshore structure”, CESLIC Report OR6, Feb. 1988 [52] Berens, A. P. and Hovey, P. W. “Characterisation of
NDE reliability” Review of progress in quantitative
[39] Kountouris I. S. and Baker M. J., “Reliability of non-
NDE, Vol. 1. 1982
destructive examination of welded joints”, CESLIC
Report OR7, Feb. 1989 [53] Temple, J. A., “The reliability of non-destructive
detection and sizing in periodic inspection of
[40] Kountouris I. S. and Baker M. J., “Defect
pressurised components”, London, IMEng, 1982,
assessment: analysis of the dimensions of defects
pp257-64
detected by ultrasonic inspection in an offshore
structure”, CESLIC Report OR8, May 1989 [54] Wolstenholme L, “A Comparison of Models for
Probability of Detection (POD) Curves”, Statistical
[41] Rodrigues, P E L B, Wong, W K and Rogerson, J. H
Report Paper No. 19, City University 1997
“Weld defect distribution in offshore structures and
their relevance to reliability studies, quality control [55] Wall, M., F. A. Wedgwood, F. A., S. Burch, S.,
and inservice inspection””, Proc. OTC, Paper 3693, “Modelling of NDT Reliability (POD) and applying
Houston, Texas, May 1980 corrections for human factors”, NDT.net, Vol.3
No.7, June 1998
[42] Dufresne J et al, “Probabilistic application of fracture
mechanics”, Int. Conf. On Fracture, 5 April, 1981 [56] Wall, Thompson, Rummel, Berens and Forli,
“Proceedings European-American Workshop on
[43] Burdekin, F. M. and Townend, P. H., “Reliability
Determination of reliability and validation methods for
Aspects of fracture on stress concentration regions
NDE”, Berlin, June 18-20 1997.
in offshore structures”, Proc. IOS’ 81 Glasgow,
Scotland, July 1981 [57] Simonen, F. A. and Khaleel M. A., “Effects of Flaw
sizing error on the reliability of vessels and piping”, J.
[44] Rogerson, J. H and Wong, W.K., “Weld defect
of Pres. Vess. Tech. ASME, Vol. 120, 1998
distribution in offshore structures and their effects on
structural reliability”, Proc. OTC, Paper 4237, [58] Jessop, T.J., et al., “Size Measurement and
Houston, Texas, 1982 Characterization of Weld Defects by Ultrasonic
Testing, Part 2”. Planar Defects in Ferritic Steels, No.
[45] Thurlbeck, S. D., Muhammed, A. and Burdekin, F.
3527/ 10/80 in TWI Report Series, The Welding
M., “A study of welding fabrication defects in two
Institute, Nov. 1980

12 Copyright©2002 by ASME
[59] Haines, N. F., Crutzen, S. and Vinche C. J., “A review
of the major PISC II round robin test results”, British
Journal of NDT, Nov. 1987
[60] Osage, D., Wirsching, P.H., and Mansour, A.E.,
“Application of Partial Safety Factors for Pressure
Containing Equipment,” 2000 ASME PVP
Conference, Seattle, WA, July, 2000
[61] Palmerg, B. “Crack growth data for the alluminium
alloys 7050-T3651 and 2024-T3”, FFA TN 1984-51,
The aeronautical research institute os Sweden,
Stockholm, 1984
[62] Yazdani N and Albrecht P., “Risk analysis of fatigue
failure of highway bridges", J. of Struc. Eng. ASCE,
Vol 113, No. 3 March 1987

13 Copyright©2002 by ASME

You might also like