Sei sulla pagina 1di 10

Session Objectives

• To understand the concept of repairs and its impact to system


reliability

• To recognize the methods of detecting trends in data


Reliability Analysis for Repairable • Why trend test is important?

System • To identify models for repairable data analysis

• To assess the reliability of repairable system

Types of Systems Non-Repairable Component

• Most systems can be categorized into two basic • Non-Repairable


types :- • Is defined as equipment after failure can no longer be repaired
• One time event or non-repairable systems
• Sensors, bearings, components of a • Limited by technical and/or cost of repair
repairable system
• Discard and replace with ‘similar’ component after failure

• Reusable or repairable systems • The component lifetime is a random variable described by a single time to
• Consist of many components failure
• Compressors, pumps, turbines
• Most complex systems fall into • Non-repairable data analysis or life data analysis is considered as one time
repairable category event, i.e. failure

3 4
Steps involved in analyzing non-repairable system/component Recurrent Data Analysis

• In contrast, for repairable system, the system may


accumulate any number of events (failures) over time
thus requires different method of analysis
Time to Distribution Goodness of Reliability
Failure Data Fitting fit test Measures • 2 random variables
• The number of failure by the age of T
• The time between successive failures Xi

• The age of the system at the i-th failure, Ti is given by

Regression Ti = X1 + X2 + X3 + … + Xi
Reliability
ReliaSoft Weibull++7 - www.ReliaSoft.com
Unreliability vs Time Plot
1.000
Unreliability

Data 1
Weibull-3P
RRX SRMMEDFM
F=6/S=0
Data Points
Unreliability Line
0.800

0.600

Unreliability, F(t)=1-R(t)
0.400

0.200

Kolmogorov-Smirnov 0.000
0.000 100.000 200.000 300.000 400.000
M

500.000
asdi Muhammad
University Technology Petronas
27/3/2012
9:51:42 PM

Time, (t)
β=0.6537, η=30.0050, γ=38.6174, ρ=0.9746

Unreliability 5 6

Stochastic point process Recurrent Data Analysis


• A stochastic process characterized by isolated events occurring • Repairable system
at instants distributed randomly over a time continuum.
What is it? • Definition: a system which, after failing to perform one or more
• Depending on the repair process, the subsequent time between
successive failures may or may not follow the distribution of its functions satisfactorily, can be restored to fully
representing the time to first failure. satisfactory performance by any method other than replacement
of the entire system
• For predictive purpose.
Why use it? Ti is cumulative time to failure
(Objective) • To model random events (e.g. random Xi is the inter-arrival time = Ti –Ti-1
variables Xi, time to ith failure)
dictionary.com

• First, you need to identify the repair process, whether it’s a


renewal process or a minimal repair process.
How to use it? x1 x2 x3 x4 x5 x6 x7 x8
• Then apply the principles governing renewal process or minimal
repair process as deemed fit. System
Ts T1 T2 T3 T4T5 T6 T7 T8 TE

7 The system can have multiple failures/event 8


Recurrent Data Analysis Recurrent Data Analysis

• If we use life data analysis on repairable system, it is as though we have 9 • Generally, LDA is more applicable to components that are replaced upon failure.
different systems where system 1 failed after t1, system 2 failed after t2
and so on… • Systems are normally made up by hundreds of different non-repairable
components.

Replaced with identical part

X1 Comp. 1 X X X
System 1
X2 LDA can be used
System 2 Comp. 2 X X for components
X3 Comp. 3
System 3 X X
X4
System 4
System X X X X X X X=failures
X9 Failures
System 9
Time to failure for system does not follow the same distribution as
components, therefore it is meaningless to fit a distribution.

9 10

Machine Failure Rate is the Sum of



Its Parts Instantaneous Failure Rate Rate of Occurrence of Failure (ROCOF)
ROCOF
A machine’s rate of
With more parts
• For repairable systems, the ‘failure rate’ is known as
Rate of Occurrence of Failure

occurrence of failure
changes as its parts ROCOF becomes ROCOF
in a Machine System

do, or do not fail. approximately constant


• Given N(t) is the counting of the number of failure occurs
in the interval (0,t],

N (t + Δt ) − N (t ) N (t , t + Δt )
Z(t) Time or Usage Age of System ROCOF ≡ ≡
Δt Δt
Green part
Hazard Rate, h(x)

replaced on
PM • ROCOF is the probability of failure in a small interval of
time divided by the length of the interval

Time or Usage Age of Part


11 12
Repair Assumptions Schematics for different repairs
• Unfortunately, as good as new assumption rarely occurs in
practice.

• Other Repair Assumptions:


• The system is in the same state after repair just as before it fails
or also known as ‘as bad as old’
• The system’s state falls between ‘as good as new’ and ‘as bad as
old’ state
• The system’s state is worse than old
• The system state is better than new
• Through defect elimination
System’s age is System’s age is in No change is
back to zero after between old and system’s age after
each repair new after repair repair

13 14

Repairable System Analysis for Repairable System

methods to analyze repairable system


Repair Categories

Perfect Repair Imperfect Repair Minimal Repair


(As Good As New) (Better Than Old, (Same As Old)
Worst Than New)

Collecting and analyzing data at


Collecting and analyzing data at Markov model for system that
the component level or failure
Homogeneous Generalized Non-Homogeneous the system level using recurrent have more than 2 states (up/
Poisson Process Renewal Process Poisson Process modes then use RBD to analyze
Applicable data analysis approach down)
(HPP) (GRP) (NHPP) system reliability
Models

16
15
Trend Test: Graphical Techniques Graphical Trend Test

• Arrange time to failure in chronological order


Cumulative Failure vs Time Cumulative Failure vs Time
12 12
y = 0.0012x + 1.4709 y = 0.1267x0.4788

• Plot cumulative failure vs. time R² = 0.963 R² = 0.9469

9 9

Cum Failures

Cum Failures
• Visually check whether or not trend exists. 6 6

• If there is no trend, same distribution for TTF is accepted and 3 3


LDA can be performed
• Otherwise, LDA is not applicable 0 0
0 2250 4500 6750 9000 0 2250 4500 6750 9000
Time (Hrs) Time(Hrs)

• Use R2 values to check the goodness of fit to straight line R2 value show a better fit compared with concave downward (decreasing) or
concave upward (increasing)

Result: No trend detected


17 18

Mann Trend Test Mann test values

• Null hypothesis: The data exhibit renewal process

• If accepted, data can be analyzed by fitting distribution to TBF data

• The test statistic is calculated by counting the number of reverse arrangement, M, among
the times between failure
• If T1, T2,…Tn are times between failures, reverse arrangement occurs whenever Ti<Tj for i<j
• Example: given TBFs as 21, 17, 48, 37, 64, 13 (days), find M.

• M value is compared against statistical value

z value can be compared


For more than 12 repairs (improvement), Against standard normal table

For degradation, M is replaced by (r(r-1)/2-M) r - number of repair


M - number of reversal
Detailed calculation of the values can be referred to

http://www.itl.nist.gov
19 20
Laplace Trend Test Test of independence

• Null hypothesis: HPP Time to Failure Are these time to


No of Failure (chronological)
• If the test is accepted, we can assume the data is IID exponentially distributed 1 50 failures
2 44 independent?
3 102
• Alternative hypothesis: NHPP 4 72
• If rejected, need to use minimal repair assumption 5 22
6 39
7 3
• General idea is to compare the mean value of TTF with the mid point of the interval 8 15
given by the following 9 197
10 88
a- start of observation time 11 49
b- end of observation time
12 128
Tj – cumulative time between failure
13 136 Use serial correlation test to check for
n’ –number of failure
Note: If the last failure occurs at the end of the observation period (i.e., tn = T),
14
15
13
8
independence.
then use n - 1 instead of n in the formula
16 25
17 22
• Check the value against Zα . Reject the hypothesis if the value outside of the Zα 18 139
critical. 19
20
210
97
21 30
22 23
• calculate L values for the following 23 13
• Failure Occurred at : 5, 40, 43, 175, 389, 712, 747, 795, 1299 and 1478 hours, with the observation ending 24 4
at 1500 hours.  25 10 No obvious cluster, independent assumption is valid
21 22

Test of independent Summary

• Mann Test – No trend detected. RP assumption is valid

Tj - Age at jth failure


n –Observed number of failures

• Laplace test – No trend, Use HPP


CM downtime

Tj - Age at jth failure


[a,b] –observation interval
n –Observed number of failures
• Serial correlation test for test of independence
Lag time

Not independent

23 24
Non-Homogenous Poisson Process (NHPP) Non-Homogenous Poisson Process (NHPP)

• NHPP assumes that the system is at the same state after repair as just before failure • The two most popular model for NHPP are:
Recall:
• Power Law model or also know as Crow/AMSAA

• NHPP Exponential law which is applicable when the Duane plot show obvious
curvature (will not be covered in the module)
Minimal repair assumption or
as bad as old after repair

• Assumption is based on the fact that most repairs involve small fraction of system’s
component, all other components remain at the same age.

• In other words, the ROCOF is assumed not to be constant.

25 26

Power Law Power Law-Graphical Method

• Power law assumes that the distribution for time to first failure (TTFF) is • λ and β can be estimated by plotting cumulative failure versus time on log-log
following Weibull distribution paper which would produce a straight line.
Recall for Weibull Distribution:
M (t ) = λt β
ln M (t ) = ln λ + β ln t
The probability density function (pdf) of the ith event given that the (i-1)th event occurred at Ti-1

• λ intercept of the line to the x axis


• For power law the expect number of failures, M(t) is given by • β the slope of the line

E[ N (t )] = M (t ) = λt β • Similar to Weibull plot


• β < 1, decreasing failure rate
• The derivative of equation above will give the failure intensity function: • β = 1, constant failure rate, HPP
• β > 1, increasing failure rate
u (t ) = M ' (t ) = λβt β −1
• Probability of having, n failure is given by Power law Poisson process

Pr[N (t ) = n]=
[M (t )]n e − M (t ) ; n = 1,2,...
n! 27 28
Power Law Power Law 30

• Alternatively, λ and β can be estimated using maximum likelihood


estimation. E[ N (t )] = M (t ) = λt β

• For a fixed observation time, T, and r failures occurring at times t1,t2,t3, β >1 β =1

…,tr β <1

Where
r- number of failure
T- observation time

29

Power Law 31
Estimates from Power Law Model

• Expected number of failure between periods (T1 and T2)


β
>1

Failure intensity function


u (t ) = M ' (t ) = λβt β −1 • Expected Reliability between periods (T1 and T2)
β
=1

• MTBF between periods (T1 and T2)


β
<1
β >1

β =1

Log-Log Scale
β <1

Linear Scale

32
Power Law Power Law-Example

● Alternatively, λ and β can be estimated using maximum • Failure data for a compressor in an offshore installation is as
likelihood estimation. shown in table below. The observation ends at the last
failure. Estimate the number of failure 6 months (180 days)
● For a fixed observation time, T, and r failures occurring at times after the observation ended.
t1,t2,t3,…,tr
Time between
Failure No failures (Days)
Answer
Where 1 272
r- number of failure 2 435 System 1

T- observation time 3 374 Beta 3.0600


Lambda (per Day) 2.5049E-9
4 45
5 9
6 36
7 84
8 40
9 11
10 67
33 34

Power Law Power Law-Time Truncated

• Special Cases: • Failure data for a compressor in an offshore installation is as shown


• Time truncated data (observation stopped at specific time) in table below. The observation ends after 1500 days. Estimate
• Occurrence of scheduled maintenance within the observation period the number of failure 6 months (180 days) after the observation
• If there are more than one system being observed. ended.

Time between Answer


Failure No failures (Days)
Beta 1.6665
1 272
Lambda (per Day) 2.95E-5
2 435
3 374
4 45
5 9
6 36

35 36
Power Law-Multiple System Generalized Renewal Process

• Failure data for a compressor in an offshore installation is as shown in table 1. The


observation ends after 1500 days. Another similar system is also observed for the same • Recall that most repairs result in
period and the failures are given in table 2. Estimate the number of failure 6 months (180 equipment being in between
days) after the observation ended.
‘AGAN’ and ‘ABAO’.

• Any failure is caused by faulty


Table 1 Table 2 component. Upon replacement,
the whole system’s age will be
Failure No Cum Time Failure No Cum Time Answer
1 433 1 272
‘newer’ than old due to new
2 680 2 707
Beta 1.08 component being replaced.
Lambda (Day) 0.0011
3 703 3 1081

• The actual age and thus system’s


reliability is estimated using GRP.

37 38

References

1. Feingold, H., and Ascher H. Repairable systems reliability: Modelling, inference, misconceptions and their
causes. New York: Marcel Dekker, 1984

2. Louit, D.M., Paccual, R. and Jardine, A.K.S., A practical procedure for the selection of time-to-failure models
based on assessment of trends in maintenance data, Reliability Engineering & System Safety, 2009: 1618-1628

3. Mettas, A. and Zhao, W. "Modeling and analysis of repairable system with general repair." Annual Reliability and
Maintainabilty Symposium. Alexandaria, Virginia: IEEE, 2005

4. Wolstenholme, L.C. Reliability modeling: A statistical approach. Chapman & Hall, 1999
Thank you
5. Yanez, M., Joglar, F. and Modarres, M. "Generalized renewal process for analysis of repairable system with
limited failure experience." Reliability Engineering and System Safety, 2002: 167-180

39

Potrebbero piacerti anche