You are on page 1of 28

Improved Noise Discrimination

for an Ammonia Tank AE Test

Martin Peacock

Specialist NDT Services


Refrigerated Ammonia Tank

 Single integrity tank: outer tank only


to contain the insulation Perlite Insulation
 Capacity 12,000 tonnes of Fill Nozzle
anhydrous ammonia
 Ammonia vapour drawn off,
compressed, cooled and returned Spray Ring
to tank via the spray ring
 Inner tank 32.3 m diameter , 21.9 m
high: 800mm annular space
 Tank integrity monitored with
periodic acoustic emission tests -33C
 Couplings welded to the outer tank
for waveguide attachment

2
AE Test History

 1986: no significant findings  Reported deterioration probably


 1997: increased AE around base due to the ice formation and
and centre roof nozzle. Repeat resulting noise (AE).
test in 4-years recommended.  A trial with an ice coated steel
 2001: further deterioration detected plate confirmed AE signals from
raising significant concerns. ice cracking were similar to
 Findings included air drawn into
those from steel.
the insulation space among other  AE from the ice may have
operational problems masked signals from defects in
 Ice <10mm thick found over the
the tank itself. At best, the AE
inner tank with an ice-perlite matrix result was uncertain.
300mm thick around the base.

3
Continued Operation of the Tank

 Tank had a history of fabrication and operational


problems, risk of stress corrosion cracking
 Plant next to a major population centre
 Shutting down for an internal inspection would bring
down the entire plant for several months
 Obtaining a reliable AE test result was essential
 A twofold approach was taken:
– Remove as much ice as possible (design and install a
dehumidification system)
– Develop an advanced AE noise discrimination method

4
AE Test Programme

 Improve all aspects of the AE test


– Waveguide design
– Loading (to maximum level and pressure)
– Instrumentation and personnel backup
– Small scale trial test six months before full test
– Review of tank operations, steps to minimise noise
– Additional waveguides to reduce sensor spacing
– Advanced Noise Discrimination

5
Wave Based Noise Discrimination

Cracks create primarily in-plane (IP) high frequency extensional waves with shear and
small low frequency flexural wave components depending on crack depth

Extraneous noise sources such as impact and friction create out of plane (OOP) signals
comprising primarily low frequency flexural and high frequency shear components.

Waveform (modal) analysis limited by use of waveguides and


resonant sensors. Development proceeded on the basis of there
being sufficient difference between crack signals and noise for
successful discrimination using a waveform classifier.

6
Pattern Recognition

 Use Vallen ‘Visual Class’ – integrated with ‘Visual AE’


analysis software and AMSY-5 instrumentation
 Separates (classifies) different wave types based on a time –
frequency based feature set
 Use waveform sets from known AE sources: supervised
learning
– The alternative, unsupervised learning means the software establishes
classes from statistical regularities of the patterns in a mixed data set
 Provides the required level of automation for processing
large numbers of waveforms
 Can take into account distortions due to use of waveguides
and resonant sensors

7
Pattern Recognition Process

EXPECTED CLASSES AE
INSTRUMENT

SELECT TRAINING DATA CLASSIFIER


PROCESSOR
DATA FILE

ANALYSIS &
DISPLAY

1 2 3 N

FEATURE EXTRACTION

BUILD FEATURE MATRIX

FEATURE SELECTION

TEST AND EXPORT


CLASSIFIER

XXXX.VCF

8
Class Information Added to Data

Operating Noise
(Classes 3 & 4)

Underlying AE
Activity
(Classes 1 & 2)

Class information available for filtering and plotting

9
Classifier Development
Perspex Strip Applied to Centre
of Plate Edge

Ice Reservoir  Record training data (sets of


different wave types)
 Test plate 1.5 x 1 m, 15mm
thick including a seam weld
 Test plate training data
Reflected IP Cracking Signal – Noise from ice cracking
Path
– Out of plane lead breaks
Surface
Mounted – In-plane lead breaks
Sensors
Waveguide – In-plane crack signals from a
Perspex strip
– Varied source – sensor
distances

10
Waveform Feature Extraction

Frequency spectra from overlapping time segments

11
Wave Feature Matrix and Classes

OOP Signals
4 3 2 1 IP Signals

Spray Noise

LF Noise

1 4 2
Time-Frequency Matrix
3
Selected Features in Red

Four classes selected including operating noise from


the tank - good separation of noise
12
Important Considerations

 Must have low sensitivity to waveform variations due to


the wave propagation distance (up to 3.5m)
– Training data with a wide range of amplitude, duration and
related values
– Minimise the effects of dispersion by using the leading part
of the waveforms
 Waveguide, sensor and preamplifier characteristics must
be consistent
 The classifier must be able to extract an ‘other’ class
– Not possible to model or record ‘genuine’ crack signals
– The classifier will fit all signals into one of the established
classes even if some do not belong to any of them
– Use ‘goodness of fit’ measures to look for outliers

13
Training Set Data

Distance (mm) 50mm 500mm 1000mm 1000 W 1350 Comment


AE Source IP OP IP OP IP OP IP OP IP OP In-Plane, Out of Plane
Pentel 0.3mm ● ● ● ● ● ● ● ● ● ~40 samples 40-66 dB

Pentel 0.5mm ● ● ● ● ● ● ● ● ● ~60 samples 60-80 dB

Ice cracking ● ● ~1350 samples 35-80 dB

Perspex cracking ● ~220 samples 37-70 dB

Tank operating noise from the trial test added later

14
Classifier Settings
Classifier V101 Test-3 TR Setting
Segments 16 Samp Freq 2MHz
Samples/segment 128 Samples 2048
Trigger offset -256 Pre-trig -512
Frequency Range 20-550kHz
Features Extracted 561

Detection
Threshold & 16 overlapping segments
Trigger Point of 128 points (~500us)

15
Classifier Results

Features 191
Classifier Results (191 Features) IP Signal OOP Noise Ops Noise WG Noise Percent
Test Plate IP Signal (29) 27 2 0 0 93.1%
Test Plate OOP Signal (32) 0 32 0 0 100.0%
Early fill and operating noise (4790) 1 64 4663 62 97.3%
Waveguide Noise (61) 0 0 1 60 98.4%

Features 306
Classifier Results (306 Features) IP Signal OOP Noise Ops Noise WG Noise Percent
Test Plate IP Signal (29) 27 2 0 0 93.1%
Test Plate OOP Signal (32) 0 32 0 0 100.0%
Early fill and operating noise (4790) 1 68 4668 54 97.4%
Waveguide Noise (61) 0 0 2 59 96.7%

Features 447
Classifier Results (447 Features) IP Signal OOP Noise Ops Noise WG Noise Percent
Test Plate IP Signal (29) 29 0 0 0 100.0%
Test Plate OOP Signal (32) 0 32 0 0 100.0%
Early fill and operating noise (4790) 0 15 4726 49 98.7%
Waveguide Noise (61) 0 0 1 60 98.4%

Including tank operating noise

16
Graphic of Class Separation

Low Freq Noise


Operating Noise

IP Signal
OOP Signal

Reliable classification of the operating and


fill noise was itself a major benefit

17
Detecting an ‘Other’ Class

Hits
Unknown Class
(Rain Showers)
Spray Ring Noise

Good Fit with Class Poor Fit with Class

Distance Ratio

18
Tank Test

 Fill from a ship (8,250 to 12,000 tonnes)


 Pressure test to 23” water gauge with ammonia vapour
 Maximum levels prior to testing <8,000t and <10” WG
V101 Level and Pressure

13000 25.0

12000 20.0
Ammonia Level (tonne)

11000 15.0

10000 10.0

Pressure spikes
due to shutting 9000 5.0

off refrigeration
during hold 8000
31/08/2008 31/08/2008 01/09/2008 01/09/2008 01/09/2008 01/09/2008 01/09/2008 02/09/2008 02/09/2008 02/09/2008
0.0

periods to 16:00 20:48 01:36 06:24 11:12 16:00 20:48 01:36 06:24 11:12

reduce noise. Time

Level Pressure

19
AE Sensor Layout

4.57
D 9 8 7 6 5 4 3 2 1 16 15 14 13 12 11 10

C 8 7 6 5 4 3 2 1 16 15 14 13 12 11 10 9
21.9

B 9 8 7 6 5 4 3 2 1 16 15 14 13 12 11 10

2.43
4.57
6m typ

6.3m typ

A 8 7 6 5 4 3 2 1 16 15 14 13 12 11 8 10 7 9

0 deg
1.5m

0.61
101.5
72 R8

71 R7 73 R9

4 Rows of 16 sensors on
70 R6 74 R10
shell: 16 sensors on roof
R15
79 80
R16
as shown
69 R5 ø 4.0m 75 R11
270 90

R14
78 77
R13
Maximum spacing 6.8m
68 R4

150kHz sensors
76 R12

67 R3 65 R1 Inlet Nozzle

66 R2 Gantry 130 deg


Spray Ring

Inner Tank Roof


Rx
AE waveguide positions
180 20
AE Data Acquisition

 80 channels: 64 on shell and 16 on the roof


 Waveguide mounted150kHz sensors
 Parametric inputs for level and pressure
 Detection threshold 35dB
 Classifier running in real time
 3,057,540 hits recorded over two days

21
Data Evaluation

 Plot different classes separately in real time for on-line evaluation


 Examine data from each sensor area after test completed
 Separate post test analysis carried out for the whole data set and
for classes 1 and 2 only (excluding operating and fill noise)

Green Trace OOP


Red Trace IP
Blue Trace Level

22
Single Channel Analysis

23
Operating and Fill Noise

The classifier was very effective in separating out the


noise from filling and tank operation (spray ring)

24
Location Plot – All Data

260

25
Location Plot – Classes 1 & 2

26
Results

 No reportable areas of AE activity requiring follow-up inspection


 Operating and fill noise accounted for 98% of recorded data
 Areas of minor AE activity revealed by excluding operating and
fill noise (classes 3 & 4) were noted for reference
 There was no indication of an ‘other’ class requiring investigation
 High confidence in the test result

27
Conclusions

 The classifier is a complex but practical tool for separating AE


signals of interest from vessel operating and other noise.
 Fill and spray noise was readily characterised and separable
from data of interest.
 Although the concern was ice noise, tank operating and fill noise
were overwhelmingly the most significant noise sources.
 In practical terms, the separation of fill and operating noise
proved the most useful aspect of employing the classifier.
 There was no evidence of AE from defects in the tank with or
without use of the classifier, even at historically high loads.
 Some minor AE sources had characteristics of mechanical
noise, probably associated with fixtures such as roof supports
and hold down brackets.

28

You might also like