- Email: [email protected]
Discrimination of ultrasonic indications from austenitic stainless-steel pipe welds
Discrimination of ultrasonic indications from austenitic stainless-steel pipe welds G.P. Singh and R.C. Manning The inspection of butt-welded stainless-steel pipe joints in nuclear power plants is routinely performed using ultrasonic non-destructive evaluation methods. Field experience, based on conventional ultrasonic signal-amplitude criteria, shows that a large number of indications are recorded. Most of these are not due to cracks, but are inherent in the geometry of the specimen. Discrimination between crack and geometric/weld (malignant versus benign) indications is principally based on operator experience, variations in signal amplitude, and the location of the reflector. Significant differences in performance exist due mainly to operator experience, fatigue, concentration, and conventional signal-amplitude evaluation criteria. In addition, the process of distinguishing the type of indications is very time consuming, as field experience and round-robin tests have shown. In response to this inspection problem, a pattern-recognition methodology has been developed to discriminate intergranular stress-corrosion cracking from geometric/weld reflectors in austenitic stainless-steel pipes. Results demonstrate that this algorithm can provide discrimination comparable to or better than those supplied by well trained operators. Preliminary results show that the pattern-recognition algorithm approach yields a better than 90% index of performance.
Keywords: ultrasonic testing, pattern recognition, stress-corrosion cracks, pipe welds The inspection of butt-welded austenitic stainless-steel pipe joints in nuclear power plants is routinely performed using ultrasonic NDE methods. Field experience and round-robin tests involving the conventional distance-amplitude curve (DAC) method for ultrasonic inspection have shown that a large number of indications are observed and documented. It has also been shown that these indications are due not only to intergranular stress-corrosion cracks (IGSCC), but also to the geometric/weld structure (Figure 1), eg suck-back, drop-through and counterbore, The manual analysis required to discriminate between geometric and actual crack indications is a very time consuming and difficult process. The discrimination is based principally on indication amplitude and arrival time (indication location), and often such analysis is subjectively dependent upon the operator. Thus, indication classification relates to the operator's skill at interpreting the sometimes subtle variations in indication amplitude as the transducer is moved manually relative to the source of the reflection. Highly experienced operators develop expertise in this regard; however, because of the extremely large number of personal, intellectual factors involved, this ability is difficult to transfer to inexperienced operators. Thus, as a consequence of experience differences as well as operator fatigue and level of concentration in general, significant variations in performance exist.
To minimize operator dependence and improve the reliability and efficiency of the discrimination process, various investigators, eg Shankar et alibi and Rose and Singhl~.31 have demonstrated the feasibility of applying digital signal-processing and pattern-recognition techniques for classifying IGSCC from geometric/weld reflector indications. In particular, Shankar used an adaptive learning network method, whereas Rose and Singh used simplified pattern-recognition algorithms. Although initial results from these algorithms looked Toe
Groove face
C°unte::;: t
) ~t~til:~ d oo opening
Fig. 1 A typical pipe weld cross-section showing two Sources of geometric/weld indications -- counterbore taper and excessive root penetration (drop-through)
0308-9126/83/060325-05 $3.00 © Butterworth Et Co (Publishers) Ltd NDT INTERNATIONAL.VOL 16. NO 6. DECEMBER1983
325
with known physical information. An algorithm was then developed that modelled this inlbrmationprocessing approach within the framework of a standard pattern-recognition methodology..
very promising, their performance on a larger database is still under evaluation. A new pattern-recognition algorithml"-Vl based on interaction between experts in ultrasonic weld evaluation and pattern-recognition theory has been developed. In this approach, all elements of patternrecognition methodology, data-acquisition protocol, training set, features and classification method were chosen differently than reported earlierl~-3l. The approach also was designed so that it offered several unique advantages: (1)
(2)
The methodology used in the development of the discrimination algorithm was as follows:
Operator experience was used to develop the discrimination algorithm. It ensured that the li~nitations in the data set had minimal effect as operator experience is based on a very large, personally internalized database. This allowed for relatively easy optimization of the decision algorithm. A single feature was used for p r i m a u discrimination: and two features were sufficient for the entire discrimination algorithm. The approach, being based on operator experience, would be understandable to them, making the potential acceptance of the technology easier.
(1)
Define the problem and determine the desired extent of discrimination (solution).
(2)
Observe methods and procedures used by "~arious operators in the field and incorporate these methods (where practical) illto the discrimination methodology.
(3)
Proceed by reasoning forward to achieve the desired goal of correct discrimination.
(4)
Using the above methodolo D and statislical pattern recognition, proceed by reasoning backward to define the desired data base.
(5)
Use feedback to optimize the data acquisition and the discrimination algorithm performance.
While watching "experienced" operators discriminate between ultrasonic indications it was observed that most of their decisions were based on the dynamics of the situation, ie by observing the general change in signal shape as the transducer was moved in various directions. Specifically, the signal varied less tbr cracks than for geometric/weld reflectors. This implied that one should acquire ultrasonic data from several locations to imitate this manual dynamic movement performed by the operator. For this reason, at any given circumferential location on the pipe, ultrasonic data from six different axial locations were acquired. The transducer was then moved in the circumferential direction, and the process repeated,
The "Flawsort" pattern-recognition algorithm approach, data-acquisition protocol, description and results along with the physical basis for the features are presented in this paper. The algorithm is presently being evaluated on a much larger database.
Approach An heuristic pattern-recognition method as it relates to the ultrasonic evaluation problem of discriminating between IGSCC and geometric/weld indications was employed, ie the approach is based on knowledge about the problem, observation, instinct, and so forth. These rules are then tested to determine performance level, and those that give acceptable performance levels form part of the final algorithm. To solve the described problem of discriminating between IGSCC and geometric indications, methods used by experienced operators to determine the sources of indications were observed. Then attempts were made to extract the most essential elements of the methods and correlate them
The data analysis approach was to use a minimum number of features in the discrimination algorithm. The data analysis showed that a more accurate classification could be achieved using one feature that was averaged (using data from several transducer locations). A preliminary decision regarding the nature of indications was made at each circumferential location. If the decision were a low probability decision, then another feature was used for discrimination purposes.
---~.u,~ ,®rRF/ / " signals /
+
,
" ~ e ~o / ~i No
~ ~ I Perform nelg~lbo~rhoodI I ~,.t o=,,f,o~o~ I .~. I scan~ r.~ o~ u ~ ~ r ~ f~ ~ ~End) ~ ~c,o=m~tl-r-u"s I - Ic'=um--ti''otl--J ~ 1
pipe
f
laverage Spatialy. Six waveforms
St~e ~ I f ~ t ~
I
c ~ o l ,,,
Search fo¢
I al~'opriate ~ RF segment I
I
I Extract Kurto~ls
l and ~ Jsignal-to-interfere~ce~
l
fe°tvres
I
I U~ln~Kurtosls obtain J o decision
|
~
I
~l~J
~"
~*~'~"
'
~
1
~
~lng ~ l - t o - i n ~ c e ~toin ~cisl~
~
Fig. 2 Flow chart for Flawsort algorithm used for classification of ultrasonic indications
326
NDT INTERNATIONAL.
DECEMBER 1983
The final decision integrated individual decisions for topological considerations. This process is referred to as the 'neighbourhood' approach, and consisted of weighting the decision at a given location so as to include the decisions at the neighbouring locations to arrive at the final decision of crack or no crack.
f j 1 / / / / /
Data acquisition protocol RF ultrasonic data were acquired from the region(s) where evaluation of ultrasonic indications was desired. Specifically, ultrasonic RF waveforms were acquired at circumferential locations 1.9 mm apart. At each circumferential location the transducer was moved in the axial direction and then skewed until the peak signal was obtained. A 20.48 $ts segment of the ultrasonic RF waveform was recorded. The segment representing the waveform corresponding to the lower one-third volume of the pipe was analysed. At the same circumferential location additional RF waveforms were acquired, each 0.76 mm apart along the axial pipe direction moving away from the centreline of weld. The transducer movements were accomplished using a Southwest Research Institute butt-weld inspection device.
Using the approach described above, a classification algorithm 'Flawsort' was developed. A flow chart for this algorithm is shown in Figure 2. During ultrasonic RF data acquisition, temporal averaging was carried out to increase the signal-to-noise ratio. A search procedure was used to locate a specific RF waveform segment; from this, two features, coefficient of kurtosis and signal-to-interference ratio (SIR), were extracted for primary and secondary discriminations respectively. The value of the coefficient of kurtosis (K), which describes the general waveshape characteristic, and the SIR of a waveform may be computed using the expression m4
(rn~)~
where 1 m 4
=
__
N
N
~, O~ --4)4
(1)
_--
and 1
rn 2 = ~ -
(At-A):
A i is the amplitude of the digitized waveform consisting of N digital points, and A is the mean amplitude value. Peak signal amplitude SIR
f Crack
b
predetermined value. If this intermediate discrimination were a high-pr0bability decision, the decision was stored. On the other hand, if the intermediate discrimination were a low-probability decision, then a secondary discrimination using the second feature, SIR, was carried out. Since the secondary discrimination also used a single feature, a simple comparison with a predetermined threshold value was sufficient. This process, which could also be described as a logic network, provided a more accurate discrimination decision. Once discrimination at specific circumferential locations was obtained, smoothing or averaging was performed using the neighbourhood approach, ie if many indications were classified as geometric/weld reflector indications in a region and only one or two isolated indications were classified as crack indications, then the final decision was that a crack was not present in that region. This was because cracks have a certain aspect ratio; and, therefore, crack indications should have a certain minimum length.
Physical basis for features
N
~= ~
/,~./f / f '~¢ // f
Fig. 3 a - - Geometrical representation of ultrasonic reflection from a corner (crack indication); b - - typical RF signal (high kurtosis)
Algorithm description
K-
El
f
Peak interference amplitude preceding the main signal
(2)
The coefficient of kurtosis feature value for the primary discrimination was compared with a predetermined threshold value (obtained from a training set). and a discrimination decision was made. The discrimination decision was further categorized as either a lowprobability or high-probability decision depending on the magnitude of the feature value in relation to the
NDT INTERNATIONAL. DECEMBER 1983
It is highly desirable to choose those features that have a physical meaning to minimize the dependence of discrimination results (derived from a small database) on purely statistical correlation. The following paragraphs, therefore, attempt to provide a brief explanation of the physical basis of both features used in the Flawsort discrimination algorithm. Consider that the surface-breaking crack plane and the inner diameter of the pipe make a right-angle plane (see Figure 3a). Ultrasonic rays striking such a rightangle edge are reflected parallel to themselves after a double reflection. Due to the special geometry, all rays travel the same distance and thus arrive in phase, constructively interfering. Therefore, the resulting ultrasonic signal is very sharp and broadband (see Figure 3b). On the other hand, ultrasonic signals from most geometric/weld reflectors are random and
327
contribution to the diffused ultrasonic signal from drop-through weld structure is random bouncing and mode conversion of sound waves (see Figure 4a). Figure 4b shows typical RF waveforms from a geometric/weld reflector. The kurtosis feature was extracted to describe the general waveshape characteristic, and has been found to be an excellent measure for determining changes in signal shape (irrespective of its amplitude) caused by the interaction of sound waves with a crack or geometric/weld reflector. The feature measures the peak of a distribution. For example, a sharply defined crack signal has a large kurtosis value compared to a geometric/weld signal, which is relatively narrowbanded (rings more). The second feature, SIR, may be explained on the basis of an ultrasonic travel path through various microstructures. This is illustrated in Figure 5 tbr both a crack (Figure 5a) and a typical geometric/weld reflector, drop-through (in figure 5b). In the case of a crack, sound travels through the base metal and heataffected zone. For a geometric/weld reflector, however, sound travels not only through the base metal and heat-affected zone but also through the weld metal (Figure 6b). The grain size and acoustic impedance differences between the base and weld metals cause different levels of scattering and reflection resulting in different SIRs for the two cases. The SIR for crack signals is much higher than that for geometric/weld reflector signals, as illustrated in Figure 6.
b Fig. 4 a - - Geometrical representation of ultrasonic reflection from a geometric reflector (geometry indication); b - - typical RF signal (low kurtosis)
~-
a
Crack
"l
Results Using the methodology outlined in the preceding paragraphs, very good classification results were obtained for discriminating between crack and geometric/weld ultrasonic indications. An index of
Base metal
Main signal
Main signal
~.._,._,~__~
b
Base metal
Fig. 5 Geometrical representation of ultrasound through various microstructures, resulting in different signal-to-interference ratios: a --crack indication; b --geometric indication
relatively narrowbanded. For example, drop-through has a random surface contour resulting in random arrival times for the various ultrasonic rays, causing random interference and a diffused signal. Another
_1 :1
•~ . = . ~ = ~
I
®
Fo, zro°
a
b
Fig. 6 Ultrasonic signals showing signal-to-interference ratios: a - - crack indications; b--geometric indications
Heat affected zone A6
- - ~ For 0 ° 270*
_
ZlO
_
- - _ _
;
_ - -
I
===,~ ~ ]
Heat affected zone
~
~ ,
~
A5 Fig. 7 Penetrant results from a 305 mm (12 in) diameter pipe containing intergranular stress-corrosion cracks. Areas 1 and 2 show where evaluation-set data were acquired
328
NDT INTERNATIONAL. DECEMBER 1983
a
b Fig. 8 Classification results obtained by using the Flawsort algorithm on two different areas. C and G represent crack and geometric/weld indications: a - - region 1, A6 pipe side (actual penetrant indication = 54.5 ram, crack prediction based on 49.5 mm); b - - region 2, A5 pipe side (actual penetrant indication = 42 rnm, crack I~e~iction based on 27 ram)
Table 1.
1 2 3 4 5 6 7 8 9 10 11 12 13 14
Classification results for evaluation data set using the Flawsort algorithm (1 in = 25.4 mm)
Pipe
Side
Data Quad. location (OD) (in)
D5/D6 D5/D6 D5/D6 D5/D6 D5/D6 D5/D6 D5/D6 ET/E8 E7/E8 E7/E8 E7/E8 E7/E8 A5/A6 A5/A6
D5 D5 D5 D5 D6 D6 D6 E7 E8 E8 E8 E8 A5 A6
2 2 2~3 4 1 1 3 1 1 3 1 2 4 3
11.26 -, 11.56 13.64 -~ 15.81 19.02 --, 20.44 37.72 --* 38.25 1.82--, 2.50 3.75--* 4.80 2 6 . 5 5 - ~ 29.10 2.90--* 3.50 2.60--, 3.50 22.45 --, 22.97 7.00 --, 7.45 71.32 -, 19.94 33.09 --, 35.11 23.14 -~ 25.84
(Indication) data length (OD) (in)
Flawsort decision based on length (in)
Flawsort decision
Penetrant test indication
Correct decision
0.30 2.17 1.42 0.53 0.68 1.05 2.55 0.60 0.90 0.52 0.45 2.62 2.02 2.70
0.30 (0.150)t (1.35)2 1.42 0.53 0.45 0.975 (0.825)t (1.35)2 0.375 0.525 0.52 0.45 2.62 (0.675)1 (0.30)2 1.875
GEOMETRY CRACK* GEOMETRY CRACK CRACK CRACK CRACK* CRACK CRACK CRACK GEOMETRY GEOMETRY CRACK* CRACK
NO YES* NO YES YES YES YES* ------NO NO YES* YES*
YES YES YES YES YES YES YES UNKNOWN UNKNOWN UNKNOWN YES YES YES YES
* Broken crack
performance better than 90% was obtained in which all crack indications were correctly classified. For example, Figure 7 shows penetrant data from a small portion of a 305 m m diameter, 22.35 m m thick, Schedule-100 austenitic stainless-steel pipe containing IGSCC. (The crack was produced by the Graphite wool technique at Ishikawajima - Harima Heavy Industries, Japan.) This is a relatively small, broken IGSCC. An evaluative set of ultrasonic RF data was acquired, from the regions marked l and 2, using a semi-automated pipe scanner system, and data were analysed using the Flawsort discrimination algorithm. The results are shown in Figure 8. In both cases a correct discrimination regarding the existence of the crack was made. In one case the crack was identified as being broken. Table l shows the evaluation results of the Flawsort algorithm on 14 reflector indications acquired from three different pipes, Of the 14 indications, seven were correctly classified as cracks and four as geometry indications: three remain unconfirmed due to the lack of reliable penetrant test results. Such parameters as reflector location, pipe identification, Flawsort algorithm decision, and verification using a penetrant test are also shown in Table 1. Currently the Flawsort algorithm is undergoing further evaluation to develop a higher degree of confidence in its performance.
has been developed. This approach appears to be very promising for providing accurate discrimination results.
References I Shankar,R., Mucciardi, A.N., Lawrie, W.E. and Stein, R.N. "Development of adaptive learning networks for pipe inspection" EPRI NP-688 (March 1978) 2 Rose, J.L and SingK G.P. 'A pattern recognition reflector classification feasibility study in the ultrasonic inspection of stainless steel pipe welds' Brit J NDT 21 No 6 (November 1979) .3 Rose, J.L. and SingK G.P. 'Stress corrosion cracking vs geometric reflector classification analysis for 304 austenitic stainless steel pipe weld specimens', Proc ofgth World Conference on NDT. Melbourne. Australia (November 1979) 4 Sing,h, G.P. 'Flaw classification in austenitic stainless steel pipes using the digital signal-processing and patternrecognition approach'. Final Report submitted to Southwest Research Institute Internal Research Panel ~(August 1981) 5 Singh,G.P. "Crack vs geometrical/weld flaw classification in austenitic stainless steel pipes'. Southwest Research Institute Internal Research Project Project No 17-9318. (October 1981) 6 Singh, G.P. nnd Mannin&R.C. 'A portable system for classification of ultrasonic indications', presented at 5th International Conference on NDE in the Nuclear Industry. San Diego. California (May 1981) 7 Singh, G.P. and Manning. R.C. "An artificial intelligence approach to ultrasonic weld evaluation'. Proc Review of Progress in Quantitative NDE. La Jolla. California (August 1982)
Authors
Conclusions A pattern-recognition methodology for classifying intergranular stress-corrosion crack ultrasonic indications and geometric/weld reflector indications
NDT INTERNATIONAL. DECEMBER 1983
Dr G.P. Singh is a Senior Research Engineer and Mr R.C. Manning is a research engineer in the Department of Research and Development, Southwest Research Institute, 6220 Culebra Road, San Antonio, TX 78284, USA.
329
Keywords: ultrasonic testing, pattern recognition, stress-corrosion cracks, pipe welds The inspection of butt-welded austenitic stainless-steel pipe joints in nuclear power plants is routinely performed using ultrasonic NDE methods. Field experience and round-robin tests involving the conventional distance-amplitude curve (DAC) method for ultrasonic inspection have shown that a large number of indications are observed and documented. It has also been shown that these indications are due not only to intergranular stress-corrosion cracks (IGSCC), but also to the geometric/weld structure (Figure 1), eg suck-back, drop-through and counterbore, The manual analysis required to discriminate between geometric and actual crack indications is a very time consuming and difficult process. The discrimination is based principally on indication amplitude and arrival time (indication location), and often such analysis is subjectively dependent upon the operator. Thus, indication classification relates to the operator's skill at interpreting the sometimes subtle variations in indication amplitude as the transducer is moved manually relative to the source of the reflection. Highly experienced operators develop expertise in this regard; however, because of the extremely large number of personal, intellectual factors involved, this ability is difficult to transfer to inexperienced operators. Thus, as a consequence of experience differences as well as operator fatigue and level of concentration in general, significant variations in performance exist.
To minimize operator dependence and improve the reliability and efficiency of the discrimination process, various investigators, eg Shankar et alibi and Rose and Singhl~.31 have demonstrated the feasibility of applying digital signal-processing and pattern-recognition techniques for classifying IGSCC from geometric/weld reflector indications. In particular, Shankar used an adaptive learning network method, whereas Rose and Singh used simplified pattern-recognition algorithms. Although initial results from these algorithms looked Toe
Groove face
C°unte::;: t
) ~t~til:~ d oo opening
Fig. 1 A typical pipe weld cross-section showing two Sources of geometric/weld indications -- counterbore taper and excessive root penetration (drop-through)
0308-9126/83/060325-05 $3.00 © Butterworth Et Co (Publishers) Ltd NDT INTERNATIONAL.VOL 16. NO 6. DECEMBER1983
325
with known physical information. An algorithm was then developed that modelled this inlbrmationprocessing approach within the framework of a standard pattern-recognition methodology..
very promising, their performance on a larger database is still under evaluation. A new pattern-recognition algorithml"-Vl based on interaction between experts in ultrasonic weld evaluation and pattern-recognition theory has been developed. In this approach, all elements of patternrecognition methodology, data-acquisition protocol, training set, features and classification method were chosen differently than reported earlierl~-3l. The approach also was designed so that it offered several unique advantages: (1)
(2)
The methodology used in the development of the discrimination algorithm was as follows:
Operator experience was used to develop the discrimination algorithm. It ensured that the li~nitations in the data set had minimal effect as operator experience is based on a very large, personally internalized database. This allowed for relatively easy optimization of the decision algorithm. A single feature was used for p r i m a u discrimination: and two features were sufficient for the entire discrimination algorithm. The approach, being based on operator experience, would be understandable to them, making the potential acceptance of the technology easier.
(1)
Define the problem and determine the desired extent of discrimination (solution).
(2)
Observe methods and procedures used by "~arious operators in the field and incorporate these methods (where practical) illto the discrimination methodology.
(3)
Proceed by reasoning forward to achieve the desired goal of correct discrimination.
(4)
Using the above methodolo D and statislical pattern recognition, proceed by reasoning backward to define the desired data base.
(5)
Use feedback to optimize the data acquisition and the discrimination algorithm performance.
While watching "experienced" operators discriminate between ultrasonic indications it was observed that most of their decisions were based on the dynamics of the situation, ie by observing the general change in signal shape as the transducer was moved in various directions. Specifically, the signal varied less tbr cracks than for geometric/weld reflectors. This implied that one should acquire ultrasonic data from several locations to imitate this manual dynamic movement performed by the operator. For this reason, at any given circumferential location on the pipe, ultrasonic data from six different axial locations were acquired. The transducer was then moved in the circumferential direction, and the process repeated,
The "Flawsort" pattern-recognition algorithm approach, data-acquisition protocol, description and results along with the physical basis for the features are presented in this paper. The algorithm is presently being evaluated on a much larger database.
Approach An heuristic pattern-recognition method as it relates to the ultrasonic evaluation problem of discriminating between IGSCC and geometric/weld indications was employed, ie the approach is based on knowledge about the problem, observation, instinct, and so forth. These rules are then tested to determine performance level, and those that give acceptable performance levels form part of the final algorithm. To solve the described problem of discriminating between IGSCC and geometric indications, methods used by experienced operators to determine the sources of indications were observed. Then attempts were made to extract the most essential elements of the methods and correlate them
The data analysis approach was to use a minimum number of features in the discrimination algorithm. The data analysis showed that a more accurate classification could be achieved using one feature that was averaged (using data from several transducer locations). A preliminary decision regarding the nature of indications was made at each circumferential location. If the decision were a low probability decision, then another feature was used for discrimination purposes.
---~.u,~ ,®rRF/ / " signals /
+
,
" ~ e ~o / ~i No
~ ~ I Perform nelg~lbo~rhoodI I ~,.t o=,,f,o~o~ I .~. I scan~ r.~ o~ u ~ ~ r ~ f~ ~ ~End) ~ ~c,o=m~tl-r-u"s I - Ic'=um--ti''otl--J ~ 1
pipe
f
laverage Spatialy. Six waveforms
St~e ~ I f ~ t ~
I
c ~ o l ,,,
Search fo¢
I al~'opriate ~ RF segment I
I
I Extract Kurto~ls
l and ~ Jsignal-to-interfere~ce~
l
fe°tvres
I
I U~ln~Kurtosls obtain J o decision
|
~
I
~l~J
~"
~*~'~"
'
~
1
~
~lng ~ l - t o - i n ~ c e ~toin ~cisl~
~
Fig. 2 Flow chart for Flawsort algorithm used for classification of ultrasonic indications
326
NDT INTERNATIONAL.
DECEMBER 1983
The final decision integrated individual decisions for topological considerations. This process is referred to as the 'neighbourhood' approach, and consisted of weighting the decision at a given location so as to include the decisions at the neighbouring locations to arrive at the final decision of crack or no crack.
f j 1 / / / / /
Data acquisition protocol RF ultrasonic data were acquired from the region(s) where evaluation of ultrasonic indications was desired. Specifically, ultrasonic RF waveforms were acquired at circumferential locations 1.9 mm apart. At each circumferential location the transducer was moved in the axial direction and then skewed until the peak signal was obtained. A 20.48 $ts segment of the ultrasonic RF waveform was recorded. The segment representing the waveform corresponding to the lower one-third volume of the pipe was analysed. At the same circumferential location additional RF waveforms were acquired, each 0.76 mm apart along the axial pipe direction moving away from the centreline of weld. The transducer movements were accomplished using a Southwest Research Institute butt-weld inspection device.
Using the approach described above, a classification algorithm 'Flawsort' was developed. A flow chart for this algorithm is shown in Figure 2. During ultrasonic RF data acquisition, temporal averaging was carried out to increase the signal-to-noise ratio. A search procedure was used to locate a specific RF waveform segment; from this, two features, coefficient of kurtosis and signal-to-interference ratio (SIR), were extracted for primary and secondary discriminations respectively. The value of the coefficient of kurtosis (K), which describes the general waveshape characteristic, and the SIR of a waveform may be computed using the expression m4
(rn~)~
where 1 m 4
=
__
N
N
~, O~ --4)4
(1)
_--
and 1
rn 2 = ~ -
(At-A):
A i is the amplitude of the digitized waveform consisting of N digital points, and A is the mean amplitude value. Peak signal amplitude SIR
f Crack
b
predetermined value. If this intermediate discrimination were a high-pr0bability decision, the decision was stored. On the other hand, if the intermediate discrimination were a low-probability decision, then a secondary discrimination using the second feature, SIR, was carried out. Since the secondary discrimination also used a single feature, a simple comparison with a predetermined threshold value was sufficient. This process, which could also be described as a logic network, provided a more accurate discrimination decision. Once discrimination at specific circumferential locations was obtained, smoothing or averaging was performed using the neighbourhood approach, ie if many indications were classified as geometric/weld reflector indications in a region and only one or two isolated indications were classified as crack indications, then the final decision was that a crack was not present in that region. This was because cracks have a certain aspect ratio; and, therefore, crack indications should have a certain minimum length.
Physical basis for features
N
~= ~
/,~./f / f '~¢ // f
Fig. 3 a - - Geometrical representation of ultrasonic reflection from a corner (crack indication); b - - typical RF signal (high kurtosis)
Algorithm description
K-
El
f
Peak interference amplitude preceding the main signal
(2)
The coefficient of kurtosis feature value for the primary discrimination was compared with a predetermined threshold value (obtained from a training set). and a discrimination decision was made. The discrimination decision was further categorized as either a lowprobability or high-probability decision depending on the magnitude of the feature value in relation to the
NDT INTERNATIONAL. DECEMBER 1983
It is highly desirable to choose those features that have a physical meaning to minimize the dependence of discrimination results (derived from a small database) on purely statistical correlation. The following paragraphs, therefore, attempt to provide a brief explanation of the physical basis of both features used in the Flawsort discrimination algorithm. Consider that the surface-breaking crack plane and the inner diameter of the pipe make a right-angle plane (see Figure 3a). Ultrasonic rays striking such a rightangle edge are reflected parallel to themselves after a double reflection. Due to the special geometry, all rays travel the same distance and thus arrive in phase, constructively interfering. Therefore, the resulting ultrasonic signal is very sharp and broadband (see Figure 3b). On the other hand, ultrasonic signals from most geometric/weld reflectors are random and
327
contribution to the diffused ultrasonic signal from drop-through weld structure is random bouncing and mode conversion of sound waves (see Figure 4a). Figure 4b shows typical RF waveforms from a geometric/weld reflector. The kurtosis feature was extracted to describe the general waveshape characteristic, and has been found to be an excellent measure for determining changes in signal shape (irrespective of its amplitude) caused by the interaction of sound waves with a crack or geometric/weld reflector. The feature measures the peak of a distribution. For example, a sharply defined crack signal has a large kurtosis value compared to a geometric/weld signal, which is relatively narrowbanded (rings more). The second feature, SIR, may be explained on the basis of an ultrasonic travel path through various microstructures. This is illustrated in Figure 5 tbr both a crack (Figure 5a) and a typical geometric/weld reflector, drop-through (in figure 5b). In the case of a crack, sound travels through the base metal and heataffected zone. For a geometric/weld reflector, however, sound travels not only through the base metal and heat-affected zone but also through the weld metal (Figure 6b). The grain size and acoustic impedance differences between the base and weld metals cause different levels of scattering and reflection resulting in different SIRs for the two cases. The SIR for crack signals is much higher than that for geometric/weld reflector signals, as illustrated in Figure 6.
b Fig. 4 a - - Geometrical representation of ultrasonic reflection from a geometric reflector (geometry indication); b - - typical RF signal (low kurtosis)
~-
a
Crack
"l
Results Using the methodology outlined in the preceding paragraphs, very good classification results were obtained for discriminating between crack and geometric/weld ultrasonic indications. An index of
Base metal
Main signal
Main signal
~.._,._,~__~
b
Base metal
Fig. 5 Geometrical representation of ultrasound through various microstructures, resulting in different signal-to-interference ratios: a --crack indication; b --geometric indication
relatively narrowbanded. For example, drop-through has a random surface contour resulting in random arrival times for the various ultrasonic rays, causing random interference and a diffused signal. Another
_1 :1
•~ . = . ~ = ~
I
®
Fo, zro°
a
b
Fig. 6 Ultrasonic signals showing signal-to-interference ratios: a - - crack indications; b--geometric indications
Heat affected zone A6
- - ~ For 0 ° 270*
_
ZlO
_
- - _ _
;
_ - -
I
===,~ ~ ]
Heat affected zone
~
~ ,
~
A5 Fig. 7 Penetrant results from a 305 mm (12 in) diameter pipe containing intergranular stress-corrosion cracks. Areas 1 and 2 show where evaluation-set data were acquired
328
NDT INTERNATIONAL. DECEMBER 1983
a
b Fig. 8 Classification results obtained by using the Flawsort algorithm on two different areas. C and G represent crack and geometric/weld indications: a - - region 1, A6 pipe side (actual penetrant indication = 54.5 ram, crack prediction based on 49.5 mm); b - - region 2, A5 pipe side (actual penetrant indication = 42 rnm, crack I~e~iction based on 27 ram)
Table 1.
1 2 3 4 5 6 7 8 9 10 11 12 13 14
Classification results for evaluation data set using the Flawsort algorithm (1 in = 25.4 mm)
Pipe
Side
Data Quad. location (OD) (in)
D5/D6 D5/D6 D5/D6 D5/D6 D5/D6 D5/D6 D5/D6 ET/E8 E7/E8 E7/E8 E7/E8 E7/E8 A5/A6 A5/A6
D5 D5 D5 D5 D6 D6 D6 E7 E8 E8 E8 E8 A5 A6
2 2 2~3 4 1 1 3 1 1 3 1 2 4 3
11.26 -, 11.56 13.64 -~ 15.81 19.02 --, 20.44 37.72 --* 38.25 1.82--, 2.50 3.75--* 4.80 2 6 . 5 5 - ~ 29.10 2.90--* 3.50 2.60--, 3.50 22.45 --, 22.97 7.00 --, 7.45 71.32 -, 19.94 33.09 --, 35.11 23.14 -~ 25.84
(Indication) data length (OD) (in)
Flawsort decision based on length (in)
Flawsort decision
Penetrant test indication
Correct decision
0.30 2.17 1.42 0.53 0.68 1.05 2.55 0.60 0.90 0.52 0.45 2.62 2.02 2.70
0.30 (0.150)t (1.35)2 1.42 0.53 0.45 0.975 (0.825)t (1.35)2 0.375 0.525 0.52 0.45 2.62 (0.675)1 (0.30)2 1.875
GEOMETRY CRACK* GEOMETRY CRACK CRACK CRACK CRACK* CRACK CRACK CRACK GEOMETRY GEOMETRY CRACK* CRACK
NO YES* NO YES YES YES YES* ------NO NO YES* YES*
YES YES YES YES YES YES YES UNKNOWN UNKNOWN UNKNOWN YES YES YES YES
* Broken crack
performance better than 90% was obtained in which all crack indications were correctly classified. For example, Figure 7 shows penetrant data from a small portion of a 305 m m diameter, 22.35 m m thick, Schedule-100 austenitic stainless-steel pipe containing IGSCC. (The crack was produced by the Graphite wool technique at Ishikawajima - Harima Heavy Industries, Japan.) This is a relatively small, broken IGSCC. An evaluative set of ultrasonic RF data was acquired, from the regions marked l and 2, using a semi-automated pipe scanner system, and data were analysed using the Flawsort discrimination algorithm. The results are shown in Figure 8. In both cases a correct discrimination regarding the existence of the crack was made. In one case the crack was identified as being broken. Table l shows the evaluation results of the Flawsort algorithm on 14 reflector indications acquired from three different pipes, Of the 14 indications, seven were correctly classified as cracks and four as geometry indications: three remain unconfirmed due to the lack of reliable penetrant test results. Such parameters as reflector location, pipe identification, Flawsort algorithm decision, and verification using a penetrant test are also shown in Table 1. Currently the Flawsort algorithm is undergoing further evaluation to develop a higher degree of confidence in its performance.
has been developed. This approach appears to be very promising for providing accurate discrimination results.
References I Shankar,R., Mucciardi, A.N., Lawrie, W.E. and Stein, R.N. "Development of adaptive learning networks for pipe inspection" EPRI NP-688 (March 1978) 2 Rose, J.L and SingK G.P. 'A pattern recognition reflector classification feasibility study in the ultrasonic inspection of stainless steel pipe welds' Brit J NDT 21 No 6 (November 1979) .3 Rose, J.L. and SingK G.P. 'Stress corrosion cracking vs geometric reflector classification analysis for 304 austenitic stainless steel pipe weld specimens', Proc ofgth World Conference on NDT. Melbourne. Australia (November 1979) 4 Sing,h, G.P. 'Flaw classification in austenitic stainless steel pipes using the digital signal-processing and patternrecognition approach'. Final Report submitted to Southwest Research Institute Internal Research Panel ~(August 1981) 5 Singh,G.P. "Crack vs geometrical/weld flaw classification in austenitic stainless steel pipes'. Southwest Research Institute Internal Research Project Project No 17-9318. (October 1981) 6 Singh, G.P. nnd Mannin&R.C. 'A portable system for classification of ultrasonic indications', presented at 5th International Conference on NDE in the Nuclear Industry. San Diego. California (May 1981) 7 Singh, G.P. and Manning. R.C. "An artificial intelligence approach to ultrasonic weld evaluation'. Proc Review of Progress in Quantitative NDE. La Jolla. California (August 1982)
Authors
Conclusions A pattern-recognition methodology for classifying intergranular stress-corrosion crack ultrasonic indications and geometric/weld reflector indications
NDT INTERNATIONAL. DECEMBER 1983
Dr G.P. Singh is a Senior Research Engineer and Mr R.C. Manning is a research engineer in the Department of Research and Development, Southwest Research Institute, 6220 Culebra Road, San Antonio, TX 78284, USA.
329