Reliability Analysis of Welds in Nuclear

Power Plant Reactor Vessels

by Pratik Tagwale

11CE31009
Under the guidance of
Prof. Baidurya Bhattacharya

Department of Civil Engineering

Indian Institute of Technology, Kharagpur
BACKGROUND
The reactor under consideration is located at
Kalpakkam, situated at 68 km south of
Chennai.
Due to the operating conditions of
temperature, humidity, pH the nuclear power
plant reactor components undergo
degradation mechanisms that affect
mechanical strength.
 In such a scenario, continuous monitoring
and inspection of critical components is the
only method with which we can ensure the
integrity of the structures and components Figure: Components of PFBR
while the plant is in service.
PROBLEM DEFINITION
The ISI of nuclear plants is far more complex than that of conventional plants
due to high radiation levels, making access to the components to be inspected
difficult.

Also it is not possible carry out frequent non destructive test due to many
constraints like inaccessibility of the area to be inspected etc.

Hence we require a time dependent reliability model to give us an estimate of

the failure probability based on periodic inspections of the reactor components
and ensure it is within safe limits
OBJECTIVE

To develop a time dependent reliability model of

components incorporating loading history, degradation
mechanism, and in service inspection data
 APPROACH
1. Develop POD (probability of detection) characteristics of given In-service Inspection
(ISI) technique.

2. Develop a methodology to update the present crack size distribution based on

single/multiple measurements from the ISI results.

3. Develop a methodology to convert the input loading history into its power spectral
density (PSD) and use it to generate multiple stationary loading histories.

4. Develop a crack growth model to predict the crack length distribution at future
instants of time due to the stationary loading histories.
Literature Survey
The primary purpose of NPP maintenance is to allow nuclear operators to use
the power plant to produce power in a safe and reliable way. Hence, adequate
maintenance is essential from the point of view of reliability of components and
equipment. From the very simple 'only fix it when it's broken' technique, Now
more advanced versions of preventive maintenance have evolved.

We will be using the techniques which give a quantitative response for the
presence of defects as we require information on the size of the defect. Hence,
Ultrasonic Testing is the most suitable method. The probability of detection
curve is then constructed for the test (Annis 2009).
The measurements for the size of defects and the characteristic POD curve is then used
for Bayesian updating of current estimated crack size distribution by incorporating
single/multiple measurements from the in-service inspection results. We can update
the present crack size distribution which we get from the manufacturer or from existing
literature using the method of Bayesian updating (Karandikar 2012)

The updated current crack size distribution along with fracture mechanics based crack
growth law parameters and loading history is then used for estimating the crack size at
the required future instant of time. From available loading time history, the power
spectral density is calculated by taking the Fourier transform of the autocorrelation
function. The power spectral density is then used to generate multiple loading histories
(Shinozuka 1972) in order to account for the randomness in actual loading.
Each simulated loading history needed to be filtered and converted into simple stress
reversals. There are many well-known cycle counting methods. It was observed that
there are many situations where unreasonable results were obtained for algorithms
other than rainflow counting (Dowling 1971). Hence it is necessary to implement
rainflow counting to the loading history because if it is not implemented, the larger
stress reversals are broken into smaller stress reversals and they tend to under estimate
the damage done to the component .

The filtered loading history is then used in a crack growth law to obtain crack size at
future instants of time. The crack growth law which we will be using is the Paris law
(Paris 1961). This crack size at a future instant of time will then be used to estimate the
reliability of the weld component at that particular instant of time.
METHODLOGY
To Calculate the Reliability at a future time instant t, we first need to predict the crack
sizes at that future time instant
Crack size at any time, AN (t) = G (t, A0 , CGL, L) (1)

A0 : Current crack size

CGL : Crack growth law

L : Loading History

We can then calculate the reliability at a future instant of time by calculating its
complement, the probability of failure.
Probability of failure Pf = P θ ∈ F = I θ q θ dθ (2)

Where Ɵ = [Ɵ1 … Ɵn] represents an uncertain state of the system with probability density
function (PDF) q

F is the failure region within the parameter space Ɵ

I is an indicator function with I (Ɵ) = 1 if Ɵϵ F and I (Ɵ) = 0 otherwise

∞
Reliability t = 1 − Probability of failure = 1 - f
ac AN aN
da (3)

ac
Reliability = f
0 AN aN
da
BAYESIAN UPDATING
In order to estimate the probability distribution of the crack size at a
later instant of time i.e. fAN(aN) we require a methodology to estimate
the present crack size distribution i.e. fA0(a0) accurately.
We might have some initial idea about the initial distribution from
tests conducted earlier or some information provided by the
manufacturer.
We can supplement this information by measurements done on the
components. We can use the initial idea about the information and then
add the information we get from measurements.
 Symbol Meaning
𝑓𝐴0 𝑎0 initial distribution of the crack size that we get from
the previous inspections or from the manufacturer

𝐴 = {𝐴1, 𝐴2, . . , 𝐴𝑛}T 𝐴 is a vector of random variables denoting the

measurements of the crack sizes taken at time t=0

X={x1,x2,..xn }T X is a vector of the observed values for 𝐴1, . . , 𝐴𝑛

𝑓 𝐴0 𝐴=𝑥 𝑎0 Updated initial distribution after getting a

measurement vector x

𝑓𝐴𝑁 𝐴=𝑥 𝑎𝑁 Updated distribution of the crack size at a future

instant of time
Bayesian Updating of initial crack distribution

𝑓𝐴 𝐴 =𝑎 𝑥 𝑓𝐴 𝑎
𝑓𝐴 0 𝐴=𝑥 𝑎0 = 0 0 0 0

𝑓𝐴 (𝑥)
Assuming Independence
𝑛
𝑓𝐴 𝐴0=𝑎0 𝑥 = 𝑓𝐴1,𝐴2,..,𝐴𝑛 𝐴0=𝑎0 𝑥1, … , 𝑥𝑛 = 𝑓𝐴 𝑖 𝐴0=𝑎0 𝑥𝑖
𝑖 1
=
. Using the total probability theorem

𝑓𝐴 1 𝐴0=𝑎0 𝑥1 = 𝑃 𝐷 ∗ 𝑓𝐴1 𝐴0=𝑎0 𝑥1

+ 𝑃(𝐷) ∗ 𝑓𝐴1 𝐴0=𝑎0 𝑥1
𝐷
𝐷

𝑥1−𝑎0 2
1
𝑓𝐴 1 𝐴0=𝑎0 𝑥1 = 𝑃𝑂𝐷 𝑎0 ∗ ∗ 𝑒− 2𝜎2 + 1 − 𝑃𝑂𝐷 𝑎0 ∗ 𝑈[0, 𝑎𝑡ℎ]
2𝜋𝜎
Calculating the Probability of Detection
Calculating the Probability of Detection
𝑃𝑂𝐷 𝑎 = 𝑃𝑟𝑜𝑏𝑎𝑏𝑖𝑙𝑖𝑡𝑦(𝑎 > 𝑎𝑡ℎ)
= 1 − 𝑃𝑟𝑜𝑏𝑎𝑏𝑖𝑙𝑖𝑡𝑦 (𝑎 < 𝑎𝑡ℎ)

𝑎 =β0 + β1 log(a)+ 𝜀 𝑤ℎ𝑒𝑟𝑒 𝜀𝑖 ~ 𝑁(0, 𝜏2)

We know that 𝑎~N(β0 + β1 log(a) , ߬ 2) 𝑎𝑡ℎ −𝛽0

Where 𝜇=
𝛽1
𝑎𝑡ℎ − (β0 + log(a)∗β1 ) 𝜏
= 1 − 𝝫( ) 𝜎=
߬ 𝛽1
log(𝑎)−μ
= 𝝫( )
𝜎
For A0~U[9,13] and x=11.25, 11.75
Algorithm to Incorporated Randomness in
Loading History

Generated Similar Calculate the crack size

Take the Loading Loading using the at the required time
History Power Spectral Density instant due to the
of the Loading History loading
Loading History
Auto-correlation Function
𝑅𝑋𝑋 𝜏 = 𝐸[𝑋 𝑡 𝑋 𝑡 + 𝜏 ]

∞
𝑅𝑋𝑋 𝜏 = −∞
𝑋 𝑡 𝑋 𝑡+𝜏 𝑑𝑡

𝑁
𝑅𝑋𝑋 𝜏 = 1 (𝑁 + 1) 𝑡=0 𝑥 𝑡 𝑥(𝑡 + 𝜏)
Auto-correlation Function
Power Spectral Density
According to the Weiner-Khintchine theorem, the Auto-correlation Function and Power
Spectral Density are Fourier transform pairs

∞
𝑆𝑋𝑋 𝜔 = 𝑅
−∞ 𝑋𝑋
𝜏 𝑒 𝑖𝜔𝜏 𝑑𝜏

𝑆𝑋𝑋 𝜔 = 𝑅𝑒( 𝑅 𝜏 𝑒 𝑖𝜔𝜏 )

𝑎𝑙𝑙 𝜏 𝑋𝑋
Power Spectral Density
Generate Random Loading Histories
𝑁
𝑋 𝑡 = 𝑘=1 [2 𝑆𝑋𝑋 𝜔𝐾 ∆𝜔]0.5 cos(𝜔𝐾′ 𝑡 + 𝜑𝐾 )

𝜔𝑢 −𝜔𝑙
∆𝜔 =
𝑁
𝜔𝐾 = 𝜔𝑙 + 𝑘 − 0.5 ∆𝜔
𝜔𝐾′ = 𝜔𝐾 + 𝛿𝜔

ωl and ωu are lower and upper limits of the frequency in the PSD
𝜑𝐾 is uniformly distributed between 0 and 2𝝅
∆𝜔 ∆𝜔
𝝳ω is uniformly distributed between [− , ]
1000 1000
Simulated Loading History
Rainflow Counting of Loading History

The rainflow-counting algorithm is used in the analysis of fatigue data in order to

reduce a spectrum of varying stress into a set of simple stress reversals.
Rainflow is a nice way to separate small, uninteresting oscillations from the large
ones, without affecting turning points.
Rainflow Counting Algorithm
Reduce the time history to a sequence of peaks and valleys such that the maximum peak or minimum
valley is at start and end of the data.
Let X denote the range under consideration; and Y, previous range adjacent to X.
1. Read the next peak or valley. If out of data, stop.
2. If there are less than three points, go to step 1. Form ranges X and Y using
the three most recent peaks and valleys that have not been discarded.
3. Compare the absolute values of ranges X and Y.
a) If X < Y, go to step 1.
b) If X ≥ Y, go to step 4.
4. Count range Y as one cycle; discard the peak and valley of Y; go to step 2.
Rainflow Counting Sample Calculation
1. Y = |A-B| and X = |B-C|. X < Y, so move to the next point,
D.
2. Y = |B-C| and X = |C-D|. X ≥ Y, so |B-C| is counted as one
cycle, and points B and C are discarded.
3. Y = |A-D| and X = |D-E|. X < Y, so move to the next point,
F.
4. Y = |D-E| and X = |E-F|. X ≥ Y, so |D-E| is counted as one
cycle, and points D and E
are discarded. Thus, the stress history of Figure 5b is
reduced to that of Figure 5c.
5. Y = |A-F| and X = |F-G|. X ≥ Y, so |A-F| is counted as one
cycle, and points A and F are discarded. Fewer than three
points are left, so the count is complete.
Crack Size Calculation using Paris Law
After performing Rainflow counting , we get a set of stress reversals.
Paris Law
𝑑𝑎
= 𝐶 (∆𝜎 𝑌 𝜋𝑎)𝑚 ∆𝑎 = 𝐶 (∆𝜎 𝑌 𝜋𝑎)𝑚 ∆𝑁
𝑑𝑁

for i=1 to m (m is number of stress reversals)

𝑎𝑗+1 = 𝑎𝑗 + ∆𝑎𝑗
where ∆𝑎𝑗 = 𝐶 (∆𝜎𝑗 𝑌 𝜋𝑎𝑗 )𝑚
end
Crack Size Distribution
Mean: 14.2321
Variance: 13.0302
Result
𝑎𝑐
𝑅𝑒𝑙𝑖𝑎𝑏𝑖𝑙𝑖𝑡𝑦 𝑡 = 0
𝑓𝐴𝑁 𝑎𝑁 𝑑𝑎

ln(𝑥)−𝑎0 2
𝑎𝑐 1 −
𝑅𝑒𝑙𝑖𝑎𝑏𝑖𝑙𝑖𝑡𝑦 𝑡 = 2047𝑑𝑎𝑦𝑠 = 0
∗ 𝑒 2𝜎
2
𝑑𝑎
2𝜋𝜎𝑥
= 99.91%
Probability of Failure=1-Reliability
=0.09%
References
1. Annis, C., Department of Defense Handbook –MIL-HDBK-1823A (2009), Nondestructive Evaluation System Reliability Assessment,
Standardization Order Desk, 700 Roberts Avenue, Philadelphia, PA 19111.
2. Karandikar, J.M., Kim, N.H., Schmitz, T.L. , Prediction of remaining useful life for fatigue-damaged structures using Bayesian inference,
Engineering Fracture Mechanics 96 (2012) 588–605
3. Hines, W.W., Montgomery, D.C., Goldsman D.M., Borror C.M., Probability and Statistics in Engineering, John Wiley & Sons; 4th Edition.
4. Lin, Y.K., and Yang, J.N. (1972). “A stochastic theory of fatigue crack propagation”, AIAA Journal, Vol. 23, No. 1 (1985), pp. 117-124.
5. Arnold, L., Stochastic Differential Equations: Theory and Applications, John Wiley and Sons, Inc., New York, 1974.
6. Lin, Y. K., Probabilistic Theory of Structural Dynamics, McGraw-Hill Book Co., 1967. Reprinted by R. E. Krieger Publishing Co.,
Melbourne, Fla., 1967
7. M Shinozuka, CM Jan, “Digital simulation of random processes and its applications”,Journal of sound and vibration 25 (1), 111-128,
1972
8. Downing, S. D., & Socie, D. F. (1982). Simplified Rainflow Counting Algorithms.International Journal of Fatigue, 4(1): 31-40
9. P.C. Paris, M.P. Gomez, and W.E. Anderson. A rational analytic theory of fatigue. The Trend in Engineering, 1961, 13: p. 9-14
Further we know that,
𝑓𝐴 𝑥 = 𝑓
𝑎0 𝐴0 𝑎0
𝑓𝐴 𝐴0=𝑎0 𝑥 𝑑𝑎0

Assuming Independence
𝑛
𝑓𝐴 𝐴0=𝑎0 𝑥 = 𝑓𝐴1,𝐴2,..,𝐴𝑛 𝐴0=𝑎0 𝑥1, … , 𝑥𝑛 = 𝑓𝐴𝑖 𝐴0=𝑎0 𝑥𝑖
𝑖 1
=
5. Using the total probability theorem

𝑓𝐴 1 𝐴0=𝑎0 𝑥1 = 𝑃 𝐷 ∗ 𝑓𝐴1 𝐴0=𝑎0 𝑥1 + 𝑃(𝐷) ∗ 𝑓𝐴1 𝐴0=𝑎0 𝑥1

𝐷 𝐷

𝑥 −𝑎 2
1 − 2𝜎
1 0
𝑓𝐴 1 𝐴0=𝑎0 𝑥1 = 𝑃𝑂𝐷 𝑎0 ∗ ∗ 𝑒 2
+ 1 − 𝑃𝑂𝐷 𝑎0 ∗ 𝛿(𝑥1 )
2𝜋𝜎

6. Hence we can calculate the updated initial distribution,

𝑓 𝐴0 𝐴=𝑥 𝑎0
Calculating the Probability of Detection
𝑃𝑂𝐷 𝑎 = 𝑃𝑟𝑜𝑏𝑎𝑏𝑖𝑙𝑖𝑡𝑦(𝑎 > 𝑎𝑡ℎ)
= 1 − 𝑃𝑟𝑜𝑏𝑎𝑏𝑖𝑙𝑖𝑡𝑦 (𝑎 < 𝑎𝑡ℎ)

𝑎 =β0 + β1 log(a)+ 𝜀 𝑤ℎ𝑒𝑟𝑒 𝜀𝑖 ~ 𝑁(0, 𝜏2)

We know that 𝑎~N(β0 + β1 log(a) , ߬ 2) 𝑎𝑡ℎ −𝛽0

Where 𝜇=
𝛽1
𝑎𝑡ℎ − (β0 + log(a)∗β1 ) 𝜏
= 1 − 𝝫( ) 𝜎=
߬ 𝛽1
log(𝑎)−μ
= 𝝫( )
𝜎