Professional Documents
Culture Documents
Repairable System Journal
Repairable System Journal
Keywords: repairable system, reliability trend test, counting failure process, homogeneous Poisson process, non-
homogeneous Poisson process, renewal process
Repairable systems reliability trend tests are reviewed, extensively tested and compared to
evaluate their effectiveness over diverse data patterns. A repairable system is often modeled as a
counting failure process. For a counting failure process, successive inter-arrival failure times will
tend to become larger (smaller) for an improving (deteriorating) system. During testing and
development of new systems, reliability trend analysis is needed to evaluate the progress of the
design development and improvement process. Often a program of testing and modification,
followed by more testing and modification, is required to achieve a desired system reliability goal.
Reliability trend tests can be an important part of this program. The objective of system reliability
trend tests is to determine whether and how the pattern of failures is significantly changing with
time. This paper reviews the following four trend tests: (1) Crow/AMSAA Test, (2) PCNT (pair-
wise comparison nonparametric test), (3) Laplace Test, and (4) Lewis-Robinson Test. These tests
are extensively tested, evaluated and compared for diverse repairable system reliability trends.
Particular emphasis focused on comparisons with low sample sizes. Simulation models for trend
tests are presented and discussed; and simulation results are summarized and compared. Based on
these comparisons, it is concluded that the Crow/AMSAA test is the most robust trend test.
1. INTRODUCTION
1
Reliability trend tests constitute a major tool during the system development or monitoring
process. They are also very helpful when reliability evaluation is needed. They can be used to
identify reliability growth by the observed data. Two general types of reliability trend tests for
repairable systems, with null (Ho) and alternative (Ha) hypotheses, as follows:
Trend Test 1.
Trend Test 2.
Ho : Renewal Process
Ha : Non-Renewal Process
For systems undergoing reliability growth testing, it is critically important to identify whether
significant improvement (i.e., real reliability growth) is occurring. System reliability growth can be
detected by observing a significant trend of increasing successive time-between-failures, i.e., system
failure inter-arrival times. For fielded systems, it is very important to detect when the system
reliability is deteriorating. Decisions for preventive maintenance and over-haul require this
information. System reliability deterioration can be detected by observing a significant trend of
decreasing successive time-between-failures. A non-homogeneous Poisson process (NHPP) is
capable of modeling these situations. If the failure intensity function, u(t), is decreasing over time,
the times between failures tend to be longer, and if it is increasing, the times between failures tend
to be shorter.
Often in system early design phase, a formalized testing program is developed to identify design
flaws and implement improvements. Reliability growth testing provides a systematic method to
conduct developmental testing, to track the progress of reliability improvement efforts and to
predict system reliability given the observed (or anticipated) rate of improvement. Reliability
growth testing has been adopted by the many industries, including defense [1], automotive [2] and
cellular telephone industries [3].
If a system in service can be repaired to "good as new" condition following each failure, then
the failure process is called a renewal process. For renewal processes, the times between failures
are independent and identically distributed. A special case of this is the Homogeneous Poisson
Process (HPP) which has independent and exponential times between failures.
Once reliability data has been collected, it is important to select an appropriate trend test. Power,
2
computational ease and simplicity of interpretation of trend test results are factors that should be
considered in the decision of selecting an appropriate trend test. Another practical problem is that
the available sample sizes are often small, making it critical to be able to evaluate the performance
of trend tests for various and differing sample sizes and select the most robust one. There are many
trend tests that will be effective when there is plentiful data. However, a truly useful trend test will
continue to be effective for smaller sample sizes.
This paper is composed of five sections. Section 2 provides formal definitions of HPP, NHPP
and Renewal Process. Section 3 introduces four trend tests. In section 4, simulation models for trend
tests are presented and discussed; simulation results are summarized and compared. Section 5
makes conclusions.
Notation
HPP Homogeneous Poisson Process
NHPP Non-homogeneous Poisson process
N(t) number of observed failures in (0, t]
u(t) failure intensity (sometimes called "instantaneous failure rate")
Λ(t) expected number of failures by time t
λ, β model parameters (λ >0, β > 0)
t development test time
Λ(t ) = ∫0 u ( s ) ds
t
(2)
Inversely, knowing Λ(t), the failure intensity at time t can be obtained as u(t)= dtd Λ (t ) .
Pr{N (t ) = n} =
[Λ (t )]n − Λ (t )
e , n = 0, 1, 2, … (3)
n!
Reliability growth testing, also known as Test-Analyze-and-Fix (TAAF) testing, involves the
testing of a system early in the development cycle when the design is immature and design changes
can be implemented more readily. At this point the design is still evolving and system reliability is
improving as design changes are made in response to observed failures. Reliability growth testing
has many advantages compared to reliability qualification and validation testing. One advantage is
that it is not necessary to wait until all design efforts have been completed to initiate a reliability
testing program. Another attractive feature of reliability growth testing is that the emphasis is on
4
reliability improvement as opposed to reliability measurement.
The concept of reliability growth testing was introduced by Duane [4] who was involved in
developmental testing of aircraft engines. Crow [5] further studied reliability growth and proposed
that the improvement in reliability can be modeled by a NHPP. He formalized his findings with the
popular Crow/AMSAA model with corresponding maximum likelihood estimators for model
parameters and goodness-of-fit tests.
E [N (t )] = λ (t )
The Crow/AMSAA reliability growth model is as follows:
β
(4)
u (t ) = λβ(t )
β −1
(5)
where N(t) = number of observed failures in (0, t]
u(t) = failure intensity
λ, β = model parameters (λ >0, β > 0)
t = development test time
For 0 < β < 1, failures during development testing occur as a NHPP with a decreasing failure
intensity. When development testing is concluded at time t, subsequent failures in actual or field
conditions at time τ, arrive in accordance with a HPP at a constant rate of u(τ;t) = λβ(t)β−1. In other
words, failure intensity is decreasing during development testing and constant thereafter. This is
logical because design fixes are being implemented to prevent or minimize the occurrences of
observed failures during the testing program. Then, improvement ceases at the conclusion of
testing because it is no longer practical or cost effective to incorporate design changes in response
to each failure.
Other reliability growth models have also been proposed by Lloyd [6], Robinson and Dietrich
[7, 8], and Crow [9 – 11].
3. TREND TESTS
During testing and development of new systems, reliability trend analysis is needed to evaluate
5
the progress of the development process. Often a program of testing and modification followed by
more testing is required in order to meet a pre-determined reliability specification [13]. Reliability
trend test can be an important part of this program.
A stochastic point process exhibits monotonic trend if FX i ( x) > FX j ( x) or FX i ( x) < FX j ( x) for
every i ≥ 1, j > i and x > 0 , where X i and X j are independent random variables [14]. If we
assume that we have a sequence of independent inter-arrival times, then this is an improving
process if FX i (x) > FX j (x) for every i ≥ 1, j > i and x > 0 . Similarly, the process is deteriorating if
the former inequality is reversed. This implies that successive inter-arrival failure times will tend to
become larger (smaller) for an improving (deteriorating) system.
The objective of system reliability trend tests is to determine whether the pattern of failures is
significantly changing with time. This can be conducted by testing a null hypothesis that the system
failure pattern is a renewal process. If this hypothesis can be rejected at some appropriate
significance level, then, it can be concluded that some level of reliability improvement or
deterioration is occurring. In practice, it can be difficult to test the renewal process null hypothesis.
It is often more convenient to test a Poisson process (HPP) null hypothesis. The danger with this
approach is that renewal processes with non-exponential failure inter-arrival times may lead to a
rejection of the null hypothesis even when there is no trend.
i =1
N
72
For large N, the test statistic is approximately distributed as a standard normal distribution
according to the central limit theorem. Therefore, reject Ho if U p > z α / 2 or U p < − z α / 2 .
7
∑T
N −1
− ( N − 1)
TN
i =1
i
2
N −1
UL= (8)
TN
12
where N= number of failures
Ti = ith failure arrival time
The rejection criteria is based on a standard normal distribution assumption for UL. It is given
by,
Reject H0 if UL > z α / 2 or UL < − z α / 2 .
The Laplace test corresponds to a Poisson process (HPP) null hypothesis. Therefore, there is the
danger that Ho is rejected when the underlying failure process is a non-exponential renewal process.
U LR =
UL
(9)
CV
where CV is the estimated coefficient of the variation of the inter-arrival times. CV can be
calculated by
CV [ X ] =
Vaˆr[ X ]
(10)
X
where X represent the variable of inter-arrival times.
Reject Ho if U LR > z α / 2 or U LR < − z α / 2 .
4. EVALUATION
9
Alpha=0.1,Beta=0.5
1.0
0.8
Probability
Crow
0.6 PCNT
Laplace
0.4
L-R
0.2
0.0
0 20 40 60 80 100
E[N(t)]
Alpha=0.1,Beta=2
1.0
0.8
Probability
Crow
0.6 PCNT
LaPlace
0.4
L-R
0.2
0.0
0 20 40 60 80 100
E[N(t)]
Alpha=0.1,E[N(t)]=10
1.0
0.8
Probabili
Crow
0.6 PCNT
LaPlace
0.4
L-R
0.2
0.0
0 1 2 3 4 5
Beta
Figure 4-3: AMSAA/Crow Model Probability Fail-to-Reject vs. β; sparse data (E[N(t)]=10)
Alpha=0.1,E[N(t)]=50
1.0
0.8
Probabilit
Crow
0.6 PCNT
0.4 LaPlace
L-R
0.2
0.0
0 1 2 3 4 5
Beta
Figure 4-4: AMSAA/Crow Model Probability Fail-to-Reject vs. β; plentiful data (E[N(t)]=50)
10
Alpha=0.1,u(T)/u(0)=0.5
1.0
0.8
Probability
Crow
0.6 PCNT
LapLace
0.4
L-R
0.2
0.0
0 50 100
E[N(t)]
Figure 4-5: Linear Model (u(T)/u(0) = .5) Probability Fail-to-Reject vs. E[N(t)]
Alpha=0.1,u(T)/u(0)=5
1.0
0.8
Probability
Crow
0.6 PCNT
LapLace
0.4 L-R
0.2
0.0
0 50 100
E[N(t)]
Alpha=0.1,E[N(t)]=10
1.00
0.80
Probability
Crow
0.60 PCNT
LapLace
0.40 ModLapLace
0.20
0.00
0 1 2 3 4 5
u(T)/u(0)
Figure 4-7: Linear Model Probability Fail-to-Reject vs. u(T)/u(0), sparse data (E[N(t)]=10)
Alpha=0.1,E[N(t)]=50
1.00
0.80
Probability
Crow
0.60 PCNT
LapLace
0.40
ModLapLace
0.20
0.00
0 1 2 3 4 5
u(T)/u(0)
Figure 4-8: Linear Model Probability Fail-to-Reject vs. u(T)/u(0), plentiful data (E[N(t)]=50)
11
Figures 4-1 and 4-5 correspond to decreasing u(t) while Figures 4-2 and 4-6 correspond to
increasing u(t). For all four of these examples, an HPP or renewal process is not appropriate
(except when β = 1 or u(T)/u(0) = 1) and a sound trend test will reject with a high probability. As
would be expected, the performance improves as more data is available. Figures 4-3 and 4-7
correspond to cases where there is sparse data available to test the trend, while Figures 4-4 and 4-8
correspond to the case where there is significant data available. For these eight graphs, it is
interesting to observe the test performance. A perfect test would have probability equal to 1-α
when β or u(T)/u(0) is one, and low everywhere else. Of course, when β or u(T)/u(0) are close to
one, it is not surprising that it is difficult to detect a trend, particularly, when data is sparse. Various
renewal processes were also simulated. Renewal processes with Weibull inter-arrival times were
selected for the simulation. These results are presented in Table 3. Figures 4-9 and 4-10 present
sample graphical output. For these analyses and graphs, β represents the Weibull distribution shape
parameter for the inter-arrival time distribution.
Alpha=0.1, Be ta=0.5
1.0
0.8
Crow
Probability
0.6 PCNT
LapLac e
0.4
L-R
0.2
0.0
0 20 40 60 80 100
E[N(t)]
1.0
0.8
Crow
Probability
0.6 P CNT
LapLac e
0.4
L-R
0.2
0.0
0 20 40 60 80 100
E[N(t)]
14
Table 1: Trend Test Results (NHPP-Linear) Table 2: Trend Test Results (NHPP-Crow)
u(T)/u(0) E[N(T)] Crow PCNT Laplace L-R Beta E[N(T)] Crow PCNT Laplace L-R
0.5 5 0.87 0.83 0.84 0.80 0.5 5 0.64 0.71 0.62 0.72
0.5 10 0.90 0.81 0.82 0.78 0.5 10 0.43 0.38 0.42 0.35
0.5 20 0.84 0.71 0.76 0.74 0.5 20 0.12 0.16 0.18 0.12
0.5 30 0.79 0.69 0.67 0.67 0.5 30 0.03 0.05 0.07 0.05
0.5 50 0.73 0.62 0.59 0.59 0.5 50 0 0.01 0.01 0.01
0.5 70 0.66 0.55 0.51 0.51 0.5 70 0 0 0 0
0.5 100 0.59 0.47 0.39 0.40 0.5 100 0 0 0 0
0.75 5 0.88 0.87 0.86 0.84 0.75 5 0.84 0.86 0.87 0.85
0.75 10 0.89 0.83 0.84 0.82 0.75 10 0.81 0.82 0.8 0.8
0.75 20 0.88 0.78 0.81 0.77 0.75 20 0.67 0.71 0.68 0.71
0.75 30 0.86 0.74 0.78 0.76 0.75 30 0.61 0.63 0.7 0.68
0.75 50 0.88 0.83 0.84 0.84 0.75 50 0.4 0.43 0.47 0.49
0.75 70 0.86 0.80 0.83 0.83 0.75 70 0.22 0.3 0.3 0.33
0.75 100 0.84 0.79 0.81 0.81 0.75 100 0.09 0.2 0.16 0.2
1.25 5 0.91 0.90 0.90 0.88 1.25 5 0.86 0.95 0.92 0.9
1.25 10 0.89 0.87 0.89 0.85 1.25 10 0.83 0.93 0.89 0.87
1.25 20 0.85 0.86 0.87 0.85 1.25 20 0.77 0.91 0.85 0.87
1.25 30 0.85 0.86 0.88 0.86 1.25 30 0.7 0.89 0.8 0.8
1.25 50 0.90 0.95 0.92 0.92 1.25 50 0.56 0.87 0.72 0.74
1.25 70 0.88 0.95 0.91 0.92 1.25 70 0.42 0.8 0.6 0.6
1.25 100 0.86 0.96 0.89 0.90 1.25 100 0.29 0.68 0.44 0.45
1.5 5 0.86 0.92 0.90 0.88 1.5 5 0.78 0.94 0.92 0.89
1.5 10 0.85 0.89 0.89 0.86 1.5 10 0.69 0.93 0.83 0.83
1.5 20 0.81 0.88 0.85 0.86 1.5 20 0.46 0.85 0.63 0.69
1.5 30 0.82 0.87 0.86 0.85 1.5 30 0.29 0.72 0.45 0.49
1.5 50 0.79 0.91 0.84 0.85 1.5 50 0.1 0.57 0.23 0.26
1.5 70 0.76 0.87 0.80 0.81 1.5 70 0.03 0.39 0.1 0.12
1.5 100 0.75 0.88 0.76 0.79 1.5 100 0.01 0.24 0.04 0.04
2 5 0.86 0.93 0.91 0.90 2 5 0.54 0.91 0.82 0.85
2 10 0.82 0.91 0.89 0.88 2 10 0.38 0.82 0.6 0.62
2 20 0.75 0.88 0.83 0.84 2 20 0.1 0.69 0.25 0.33
2 30 0.73 0.86 0.79 0.80 2 30 0.01 0.46 0.07 0.1
2 50 0.64 0.82 0.68 0.69 2 50 0 0.25 0 0.01
2 70 0.56 0.74 0.58 0.59 2 70 0 0.11 0 0
2 100 0.41 0.60 0.40 0.41 2 100 0 0.03 0 0
3 5 0.79 0.93 0.91 0.89 3 5 0.22 0.85 0.55 0.71
3 10 0.73 0.92 0.85 0.86 3 10 0.05 0.69 0.17 0.32
3 20 0.63 0.85 0.72 0.74 3 20 0 0.4 0.01 0.04
3 30 0.52 0.79 0.62 0.63 3 30 0 0.22 0 0.01
3 50 0.39 0.66 0.42 0.43 3 50 0 0.06 0 0
3 70 0.23 0.48 0.24 0.27 3 70 0 0.03 0 0
3 100 0.14 0.31 0.12 0.12 3 100 0 0 0 0
5 5 0.74 0.93 0.91 0.89 5 5 0.03 0.76 0.21 0.54
5 10 0.62 0.90 0.77 0.80 5 10 0 0.59 0.02 0.16
5 20 0.46 0.80 0.59 0.63 5 20 0 0.31 0 0.01
5 30 0.30 0.67 0.40 0.42 5 30 0 0.1 0 0
5 50 0.13 0.44 0.17 0.19 5 50 0 0.02 0 0
5 70 0.10 0.28 0.09 0.09 5 70 0 0.01 0 0
5 100 0.01 0.12 0.01 0.01 5 100 0 0 0 0
15
Table 3: Renewal Process Trend Tests −Weibull inter-arrival times
Shape, γ
Table 4: HPP Trend Test Results
E[N(T)] Crow PCNT Laplace L-R E[N(T)] Crow PCNT Laplace L-R
0.5 5 0.55 0.88 0.43 0.81 5 0.91 0.85 0.87 0.86
0.5 10 0.56 0.89 0.48 0.79 Linear 10 0.89 0.79 0.85 0.80
0.5 20 0.53 0.88 0.5 0.79 Model 20 0.87 0.80 0.84 0.80
0.5 30 0.52 0.9 0.48 0.81 u(T)/u(0) 30 0.86 0.78 0.84 0.81
0.5 50 0.54 0.89 0.52 0.82 1 50 0.91 0.89 0.91 0.91
0.5 70 0.51 0.88 0.5 0.84 70 0.92 0.94 0.95 0.95
0.5 100 0.52 0.9 0.52 0.86 100 0.93 0.94 0.95 0.95
0.75 5 0.83 0.92 0.74 0.87 5 0.90 0.84 0.86 0.87
0.75 10 0.77 0.89 0.7 0.82 Crow 10 0.88 0.80 0.85 0.80
0.75 20 0.78 0.9 0.74 0.85 Model 20 0.87 0.79 0.83 0.81
0.75 30 0.77 0.89 0.75 0.87 Beta=1 30 0.84 0.76 0.83 0.79
0.75 50 0.78 0.9 0.72 0.87 50 0.91 0.88 0.89 0.88
0.75 70 0.79 0.91 0.77 0.88 70 0.91 0.92 0.93 0.94
0.75 100 0.79 0.9 0.77 0.88 100 0.94 0.95 0.95 0.96
1.25 5 0.95 0.89 0.93 0.86 5 0.90 0.89 0.84 0.86
1.25 10 0.95 0.89 0.94 0.87 10 0.89 0.90 0.87 0.86
1.25 20 0.96 0.91 0.96 0.89 Weibull 20 0.90 0.88 0.87 0.85
1.25 30 0.95 0.91 0.95 0.9 Model 30 0.90 0.90 0.87 0.87
1.25 50 0.96 0.9 0.96 0.89 Shape 50 0.91 0.89 0.87 0.87
1.25 70 0.94 0.89 0.94 0.87 γ=1 70 0.90 0.90 0.89 0.88
1.25 100 0.94 0.9 0.93 0.9 100 0.90 0.90 0.88 0.88
1.5 5 0.96 0.91 0.97 0.87
1.5 10 0.97 0.89 0.97 0.87
1.5 20 0.98 0.9 0.98 0.89
1.5 30 0.98 0.91 0.99 0.89
1.5 50 0.98 0.9 0.99 0.9
1.5 70 0.98 0.91 0.97 0.89
1.5 100 0.97 0.9 0.97 0.9
2 5 0.99 0.91 1 0.88
2 10 0.99 0.9 1 0.89
2 20 0.99 0.9 1 0.88
2 30 0.99 0.91 1 0.91
2 50 1 0.9 1 0.89
2 70 1 0.89 1 0.9
2 100 1 0.9 1 0.91
3 5 1 0.9 1 0.86
3 10 1 0.92 1 0.9
3 20 1 0.91 1 0.88
3 30 1 0.89 1 0.88
3 50 1 0.91 1 0.91
3 70 1 0.89 1 0.89
3 100 1 0.9 1 0.9
5 5 1 0.91 1 0.88
5 10 1 0.89 1 0.88
5 20 1 0.9 1 0.87
5 30 1 0.9 1 0.89
5 50 1 0.89 1 0.89
5 70 1 0.9 1 0.9
5 100 1 0.9 1 0.9
16
REFERENCES
18