Corrosion Problems Quantified With Gumbel Lower Distribution

Corrosion Problems Quantified
With Gumbel Lower Distribution

Abstract: Several case studies show how to separate general corrosion from accelerated
corrosion and how to predict end of useful life of products.
Paul Barringer, P.E.

Barringer & Associates, Inc.
P.O. Box 3985
Humble, TX 77347-3985
Phone: 281-852-6810
FAX: 281-852-3749
Email: hpaul@barringer1.com
Web: http://www.barringer1.com
© Barringer & Associates, Inc. 2007 1
Gumbel Upper or Gumbel Lower?

• The Gumbel upper distribution is used when
you have BIG numbers. It’s best know for
flood data (you only record the deepest
[largest] stream gage reading for a single year).
• The Gumbel lower distribution is used when
you have LITTLE numbers. It’s used where
you’ve only recorded the thinnest [smallest]
wall in a single corrosion area.
The Gumbel Smallest Extreme Value is considered a model for a system having n elements in a series and
where the failure distributions of components are reasonably uniform and similar (See British Standard BS 5760).
1
The Weibull distribution straight line equation
ln⎜⎛ ln⎛⎜ ⎞⎞
⎟ ⎟ β ⋅ ln( t) − β ⋅ ln( η )
1
⎝ ⎝ 1 − F( t) ⎠ ⎠
Both are also What’s The Math Difference?

known as the The Gumbel largest extreme value CDF is: The Gumbel smallest extreme value CDF is:
− ( t − ξ) t− ξ
double δ δ
−e −e
exponential F( t ) e F( t ) 1−e
Rearanging the equations to read Rearanging the equations to read
− ( t − ξ) − ( t − ξ) t− ξ t− ξ Observations:
δ δ δ δ
−e 1 1 e −e 1 1 e
F( t ) e Or e 1 − F( t ) e Or e
− ( t − ξ) F( t ) t− ξ 1 − F( t ) Same Y-axis
δ δ
e e
e e
Weibull has
Taking the log of both sides you get: Taking the log of both sides you get:
log X-axis
− ( t − ξ) t −ξ
ln⎛⎜
1 ⎞
ln⎛⎜ ⎞ e
δ 1 δ
⎟ e ⎟ Gumbel has
⎝ F( t) ⎠ ⎝ 1 − F( t) ⎠ uniform
Again, taking the log of both sides you get: Again, taking the log of both sides you get: X-axis
−( t − ξ) −t ξ t−ξ ξ
ln⎜⎛ ln⎛⎜ ⎞⎞ ln⎛⎜ ln⎛⎜ ⎞⎞
1 1 t
⎟⎟ + ⎟⎟ −
⎝ ⎝ F( t) ⎠ ⎠ δ δ δ ⎝ ⎝ 1 − F( t) ⎠ ⎠ δ δ δ
Y = mX + b Y = mX + b
ξ is a scale factor
ξ − δ⋅ ln⎜⎛ ln⎛⎜
1 ⎞⎞
ξ + δ⋅ ln⎛⎜ ln⎛⎜ ⎞⎞
1
t ⎟⎟ t ⎟⎟
δ is a shape factor
⎝ ⎝ F( t) ⎠ ⎠ ⎝ ⎝ 1 − F( t) ⎠ ⎠
For Monte Carlo modeling: For Monte Carlo modeling: Small δ steep
t ξ − δ⋅ ln( −ln( a_random_n o) ) t ξ + δ⋅ ln( −ln( 1 − a_random_n o) )lines for G- & G+
distributions
Problem 1: Heat Exchanger Thin Tubes?
• We have a shell & tube heat exchanger

• Process fluids are inside the tubes and the
tubes are loosing wall thickness with use
• Outside the tubes are cooling water
• Periodic inspections have recorded the
minimum wall thickness in each tube
selected randomly. We have only one wall
thickness for each tube inspected.
2
What’s The Issue? How To Resolve?
• Heat exchanger is 17 years old—460 tubes

• At turnaround, eddy current wall thickness
inspection occurred—We’re worried!
• Did an IRIS inspection on 10% of tubes—Now
we’re more worried—what does the data say?
• Retube NOW at 17 years with T/A delays?
Retube next turnaround in 3 years at 20 years?
Retube at 2nd turnaround in 6 years at 23 years)?
© Barringer & Associates, Inc. 2007
Time Issues 5
What Are Cost Consequences?

• Failure $ is dependent on outside temperatures:
– Summer failure = $750,000 lost margins & retube
– Fall failure = $500,000 lost margins & retube
– Winter failure = $100,000 lost margins & retube
– Spring failure = $250,000 lost margins & retube
• Another key issue is environmental impact
along with the cost issues if failure occurs
Murphy says: Big Money Issues Will Prevail
3
Why Did They Inspect?
• Rule of thumb for this facility-
– Inspect tubes if wall thickness has been
reduced by 1/3, i.e. from 0.083” to 0.055”
– Consider retubing heat exchangers when tube
wall thickness has been reduced to ½ of
original wall thickness, i.e. when wall thickness
has been reduced from 0.083” to 0.0415”
• This exchanger has environmental concerns
Eddy Current vs IRIS Inspection

• Eddy current inspection is the usual quick
and inexpensive inspection of each tube—
minimum wall is reported for each tube
• IRIS inspection is a more detailed and more
expensive inspection with a rotating head
ultrasonic tool—minimum wall is reported
for each tube and tube ID’s must be very
clean for an accurate IRIS inspection.
4
What Did IRIS Inspection Find?
• The minimum wall thickness report shows:
Rule of thumb triggers Wall*qty
inspection at 0.050” 0.050*1 0.063*9
0.055*1 0.064*9
Wall thickness
0.056*2 0.065*4
measured
0.058*2 0.066*5
in inches
0.059*1 0.067*2
0.061*6 0.069*4
• Minimum allowed wall thickness is 0.036”
for structural integrity.
Stacks Of Data—Use Sherwin’s

Inspection Option
ges
(low) Probability Of
a
ure
Occurrence (high)
Use top l
Fai
of stack for
regression
Discovery
Benign failure Age/Thickness
occurred here?
Benign failure
discovered here
Wall Thickness Discovered At Inspection

We have stacks of data from the heat exchanger inspection because
the IRIS data have been rounded to three significant digits.
5
Competing Models:
Weibull? or Gumbel Distributions?
Weibull Distribution
with rank regression
& inspection option
Data stacks from

course measurements
∴use inspection option
for regression
Small risk of wall thickness R= Coefficient

© Barringer & Associates, Inc. 2007 of regression 11
less than min allowed ccc= critical correlation coefficient
Competing Models:
Weibull or Gumbel Distributions?
Gumbel- Distribution
with rank regression
Bigger than for

Weibull distribution
∴ use Gumble-
Higher risk of wall thickness less than R2= (Coefficient

© Barringer & Associates, Inc. 2007 of regression)2 12
min allowed ∴more conservative (ccc) 2= (critical correlation coefficient)2
6
PDF Curves
~2*x Note the Gumbel- distribution says to

x expect more occurrences with thinner walls
PDF Details
(0.050459”, 3.933565)
( t − ξ) 100 t− ξ
⎛ t− ξ ⎞ δ ⎛ t− ξ ⎞ δ
⎛ 1 ⎞ ⋅ e⎜⎝ δ
⎟ −e
⎠ ⎛ 1 ⎞ ⋅ e⎜⎝ δ
⎟ −e
⎠
⎜ ⎟ ⎜ ⎟ Gumbel Lower PDF
⎝ δ⎠ ⎝ δ⎠
2
β β
⎛ t− t0 ⎞ 50 ⎛ t− t0 ⎞
−⎜ ⎟ −⎜ ⎟
⎡⎛ β ⎞ ⋅ t − t β − 1⎤ ⋅ e ⎝
⎡⎛ β ⎞ ⋅ t − t β − 1⎤ ⋅ e ⎝
⎢⎜ β ⎟ (
0 ⎥ ) η ⎠
⎢⎜ β ⎟ (
0 ⎥ ) η ⎠
⎣⎝ η ⎠ ⎦ ⎣⎝ η ⎠ ⎦
0 0
0.04 0.06 0.04 0.05
t t
Weibull PDF
7
Why Gumbel Lower Distribution?
• The SMALLEST value is recorded for each

tube thickness which motivates use of the
Gumbel smallest distribution. Just as for
flood data (the largest yearly value) motivates
the use of the Gumbel largest distribution.
• The Gumbel smallest distribution is a better
curve fit and shows greater % potential
failure than Weibull, thus more conservative.
Here’s Where We Are At Year 17. Can We Make Year 20?

Heat Exchanger IRIS Inspection Data
99
95 G-/rr/insp1
90 Year 17 Area is 1% high
80 by 0.01” wide
70
60
Consider retube if less than 0.0415”
Inspect if less than 0.055”
50
Occurrences CDF %
Structural minimum is 0.036”
40
30
20
10 Parameters:
ξ =0.06427 Location
5 Slope/Shape
2 δ Xi Del r^2 n/s

0.06427 0.00316 0.989 46/0
Area is 1% high
by 0.01” wide 1
note the .03 .04 .05 .06 .07 .08 .09 .1 .11
magnification! Tube Wall Thickness (inches)
Small δ steep line slope
© Barringer & Associates, Inc. 2007
Large δ flat line slope. 16
8
General Corrosion
This becomes a t=? t=9 t=6 t=3yrs Start = datum
critical value!
General Deterioration 63.2%

Note Parallel Lines
Probability of
Occurrence
Don’t exceed this

probability of thin wall
Low probability
of thin wall below
Wall Thickness
minimum! © Barringer & Associates, Inc. 2007 17
min
General Corrosion Trend Line

Wall Thickness , ξ
Characteristic
Critical Value, ξ,
For Wall Thickness
Time End of life!

An easy decision.
9
Accelerated Corrosion
t=9 t=6 t=3 start
99.9%
General Deterioration
Probability of
Occurrence
Accelerated
Deterioration
Breaks The
! Min Wall
Don’t exceed this Limits!
probability of thin wall
You must know when to

accept the risk of failure
and when to accept the
risk of failure!
Wall Thickness
$Risk = pof*$Consequence min
Accelerated Corrosion Trend Line

Wall loss from general corrosion
Specified Risk—say 0.1%
Wall loss from both general + accelerated corrosion

Wall Thickness At A
Wall loss from accelerated corrosion
Minimum Wall Thickness At .

Acceptable Risk Level
Time End of life!
Difficult decision.
10
You Must Know Wall Thickness At Time Zero
Heat Exchanger IRIS Inspection Data
99.9
99
G-/rr/insp1
95 Year 17 Assumes new tube
90
80 with tmin = 0.083”
70
60 and tmax = 0.101”
50
Occurrence CDF %
40 Year ?? for ~6*σ = 99.8%
30
20
10 Min Allowed Wall = 0.036”

5 Year 0
Note the flatter slope
2
with larger δ means
1 more wall thk. scatter!
.5 ξ = Xi δ = Del r^2 n/s
0.06427 0.0031573 0.989 46/0
.2 0.09706 0.0020362
.1
.03 .04 .05 .06 .07 .08 .09 .1 .11
Tube Wall Thickness (inches)
Typical Corrosion rate = (0.09706-0.06427)/17 = ~0.002”/yr
Wall Thickness @ 99.9%

0.101 @ 0 years
0.07034 @ 17 years
0.06496 @ 20 years
0.05956 @ 23 years
Data needed for construction

of trendlines on next page
with “as new” slopes.
11
Heat Exchanger Construction Lines
99.9
99
G-/rr/insp1 0.06496 0.07037 0.101
95
90
80 General Corrosion
70
60
50
Occurrence CDF %
40 Year 20
30
20
Min Allowed Wall = 0.036”

Year 17
10
5 As New Slope Year 0
2
1 Accelerated
Corrosion
.5 Effects
.2
.1
Year 23
.03 .04 .05 .06 .07 .08 .09 .1 .11
Tube Wall Thickness (inches)
0.03531 0.05237 0.083
0.04246
Year 20 Forecasted Line: ξ = 0.05848, δ = 0.0033541 with 0.1228% occurrence at 0.036” wall.
Wall Thickness at 0.1% Risk vs Time
0.083
0.05237
Y=0.083-0.0018017t
0.04246
Y=0.083-0.0023847t
0.03531 @ 20 years
17
0.02815 @ 23 years General + Accelerate

Corrosion Rate
12
Retube Or Not Retube Now?
• At year 20 (next turnaround) the minimum
wall thickness will decline to just under 0.036”
• The risk for falling below 0.036” min wall is
0.1228% Time & Money Issues Converge
• $risk = (prob. of failure)*$Consequence, ∴

$risk exposure = 0.1228%*$750,000 = $921
• ∴ take the risk for running 3 more years
—Do not retube now. Run to TA at yr 20.
Tube Exchanger Summary

• Avoided the recently discovered and recently
expected turnaround delay for accelerated delivery of
heat exchanger ($750,000 expenditure avoided)
based on use of one day analysis of data.
• Pressing on toward the next turnaround three years
into the future
• At year 20, install a new tube bundle.
• What’s the risk for continuing to year 23?
0.91%*$750,000 = $6,825—if risk adverse, reject.
If risk accepting—maybe, but very doubtful.
13
Problem 2: Column Corrosion
• A column is rapidly loosing wall thickness.
• Fluids/gasses within the column are violent.
• Frequent Inspections—data is all over the map!
• Loss of containment will impact personnel and
environment issues with big $’s
• What should we do:
--Run?—if so, for how long?
--Shut down?—if so, how to persuade the
management team?
Developed Outer Surface Of Tower

Circumference
Inspection
Grid
Over “Bad”
Spots
Height
Data collection
on the grid
will contain both
good walls and
bad walls!
14
Raw Data UT Inspections
Thick —
Rapid ignore!
Deterioration
In Wall
Thickness Thin —
worry!
Remaining Wall Thickness (Mils/10)
Truncated Data—Thin Data Only
Remaining Wall Thickness (Mils/10)

15
End Points For Corrosion Curve
UT Wall Thickness Construction Lines 51
General Corros. 33.09
@ 0.1%
Days Thickness 32.3
0 49
906 31.09
966 25.17
1105 25.56
1127 25.17
Gen + Accel Cor.
Gen + Accel Cor. @ 99.9%
@ 0.1% Days Thickness
Days Thickness 0 51
0 49 906 33.09
906 25.56 966 32.3
966 23.61 25.17 1105 27.56
1105 19.26 1127 27.17
1127 19.53 31.09
25.56 49
25.17
End Of Life Clearly Shown
Y=49.065-0.02128*X General Corrosion
General + Accelerated Corrosion
1178
1460
End Of Life 949 1176
16
Summary
• ASME minimum wall was violated at 949 days
• API fitness for service will be violated at 1176 days
and we are 1127 days into service
• Plan an immediate orderly shutdown for replacement
• Outage + planned replacement =$10,000,000
• Emergency outage + emergency replacement =
$20,000,000 because of safety hazards
• Risk is too high! 0.1%*$10,000,000 = $10,000 and
climbing toward $20,000,000. Take action now!
Now, For Grins

• Consider the Gumbel larger distribution
• Houston flood
• Aircraft gust loads
• Space shuttle rocket motor O-ring burns
Discussions about the
Gumbel lower distribution
always raise questions
about the Gumbel upper
distribution
17
It Rained A Little
On June 9, 2001—23 Inches!
Cars are
submerged
on US 59 © Barringer & Associates, Inc. 2007 35
highway!
Peak Annual Stream Flows-Gage Height (feet)

USGS 08074000 Buffalo Bayou at Houston, Texas
99.9
G+/rr
Depth for 100 yr flood

comes from the return
99
period, RP = 1/(1-p).
When RP = 100 years,
Occurrence CDF (%)
then p = 99%
95
Forecasted One Hundred
Year Flood Gage Height
June 9, 2001
90
80
70
60
50
40
30
20
10 Xi Del r^2 n/s
5
21 17.2 5.99 0.978 67/0
.2.5.1
44.76
0 10 20 30 40 50 60 70 80
Peak Gage Height (feet)
The flood was bad but not the worst © Barringer & Associates, Inc. 2007
recorded near downtown Houston! 36
18
Assumed Houston Flood Cost In June 2001
5
Flood Cost Estimates In June 2002
Assumed Flood Cost US ($ Billion)

4
100 year flood

2
will be a 2X
$ problem
June 9, 2001
1
0
10 20 30 40 50
Gage Height (feet)

Aircraft Positive Gust Loads
19
Space Shuttle Burned O-Rings
Failures
STS 51-C Field Joint Only
3 ●
Failures
Data
And
61A 53*3
● Successes
Number of
2
57*1
Incidents
58*1 Data
41B
●●
41C
● ●41D
63*1 53*3
1 ●STS-2
61C
70*2 57*1
75*2 58*1
0 63*1
45o 50o 55o 60o 65o 70o 75o 80o
Calculated Joint Temperature, oF -66*1
STS 51-C Field Joint -67*3
3 ● -68*1
-69*1
●
61A
70*2
Number of
2
Incidents
-70*2
41B
-72*1
1 ●●
41C
● ●41D
●STS-2 -73*1
Flights 61C
with no
incidents
75*2
-76*2
0
45o 50o 55o 60o 65o
●●
●●●●
● 70o
●●
● ●● 75●
o
●●80
●●o -78*1
Calculated Joint Temperature, oF -79*1
Source: Engineering Ethics, Gail D. Baura, -80*1
Elsevier, ISBN 13:978-0-088531-2, 2006, Page 73. -81*1
Data from the Rogers Commission 1986
Good Practice Advice—Watch Out!

• Gumbel upper & lower distributions allow the use
of negative numbers on the X-axis
• When using suspensions (as a – sign) make sure
you turn on display of the suspensions (under
magnifying glass) so you can view they are in the
correct locations AND (under the Method icon)
make sure to turn the negative sign to indicate
suspension!
• Else, you’ll get misleading results.
20
Which Plot?
Suspended data
shown on plot as >
Poor curve fit

Gumble Upper Slightly Better-But

Not Every Data Fits A Plot!
Failures were resolved by rocket joint/O-ring redesign
Failures demonstrated to exist
© Barringer & Associates, Inc. 2007 If you fail to turn on “-” is a suspension 42
Better but not good curve fit you will conclude this is a good fit!!
21
Gumbel Upper Summary
• Works well when you have the largest
recorded data such as flood data, fatigue
data, etc.
• Watch for traps with suspensions when used
without good practices can result in bad
conclusions.
• If Weibull, Lognormal, etc. don’t work then
don’t expect automatic success with all data
by use of the Gumbel upper distribution.
Want More Details?

• Got to http://www.barringer1.com/problem.htm
• Look at WinSMITH Weibull software (which also includes Gumbel large
and small distributions)
• See biographies at http://www.barringer1.com of Dr. Weibull and Dr.
Abernethy who is the world’s leading expert in Weibull analysis
• Dr. Weibull got many of his ideas on extreme values while working at
Bofors Steel in Sweden—you can see Bofors antiaircraft guns at the
Museum of the Pacific in Fredricksburg, TX.
• See Gumbel, E. J., Statistics of Extremes, Columbia University Press, New
York, 1958
• See Statistical Theory of Extreme Values And Some Practical
Applications, A Series of Lectures, PB 175818, 12 Feb 1954 by Emil J.
Gumbel, National Bureau of Standards, U.S. Dept of Commerce, NTIS
• See A New Method Of Analyzing Extreme-Value Data, NACA
TN 3053, Jan 1954, U.S. Dept of Commerce, NTIS, by
Julius Lieblein National Bureau of Standards
22

Corrosion Problems Quantified With Gumbel Lower Distribution

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Corrosion Problems Quantified With Gumbel Lower Distribution

Uploaded by

Copyright:

Available Formats

Corrosion Problems Quantified

With Gumbel Lower Distribution

Paul Barringer, P.E.

© Barringer & Associates, Inc. 2007 1

Gumbel Upper or Gumbel Lower?

Both are also What’s The Math Difference?

Problem 1: Heat Exchanger Thin Tubes?

• We have a shell & tube heat exchanger

• Heat exchanger is 17 years old—460 tubes

What Are Cost Consequences?

Eddy Current vs IRIS Inspection

Stacks Of Data—Use Sherwin’s

Wall Thickness Discovered At Inspection

Data stacks from

Small risk of wall thickness R= Coefficient

Bigger than for

Higher risk of wall thickness less than R2= (Coefficient

~2*x Note the Gumbel- distribution says to

© Barringer & Associates, Inc. 2007 13

© Barringer & Associates, Inc. 2007 14

• The SMALLEST value is recorded for each

Here’s Where We Are At Year 17. Can We Make Year 20?

Inspect if less than 0.055”

Structural minimum is 0.036”

2 δ Xi Del r^2 n/s

General Deterioration 63.2%

Don’t exceed this

General Corrosion Trend Line

Time End of life!

You must know when to

Accelerated Corrosion Trend Line

Wall loss from both general + accelerated corrosion

Wall loss from accelerated corrosion

Minimum Wall Thickness At .

10 Min Allowed Wall = 0.036”

Wall Thickness @ 99.9%

Data needed for construction

© Barringer & Associates, Inc. 2007 22

Min Allowed Wall = 0.036”

Wall Thickness at 0.1% Risk vs Time

0.02815 @ 23 years General + Accelerate

• $risk = (prob. of failure)*$Consequence, ∴

Tube Exchanger Summary

Developed Outer Surface Of Tower

© Barringer & Associates, Inc. 2007 28

Remaining Wall Thickness (Mils/10)

© Barringer & Associates, Inc. 2007 29

Truncated Data—Thin Data Only

Remaining Wall Thickness (Mils/10)

End Of Life Clearly Shown

Y=49.065-0.02128*X General Corrosion

General + Accelerated Corrosion

End Of Life 949 1176

© Barringer & Associates, Inc. 2007 32

Now, For Grins

Peak Annual Stream Flows-Gage Height (feet)

Depth for 100 yr flood

Flood Cost Estimates In June 2002

Assumed Flood Cost US ($ Billion)

100 year flood

Gage Height (feet)

Aircraft Positive Gust Loads

© Barringer & Associates, Inc. 2007 38