Professional Documents
Culture Documents
Abstract - The two-parameter Weibull distribution is the II. MAXIMUM LIKELIHOOD METHOD
most widely used distribution in reliability engineering.
Parameter estimation is a key issue to apply the Weibull to In this paper, the probability density function (PDF) of
practical engineering. This paper aims to compare the two parameters Weibull density function is defined as
performance of the maximum parameter estimation method /
(MLE) and the moment parameter method. It firstly ; 0, 0, 0 (1)
investigates some mathematical properties such as solution The is the scale parameter and the is the shape
uniqueness, estimator’s equivariance and the confidence parameter.The MLE assumes the most likely parameter as
interval, that are important to practice. Later on their
the desired estimator. This principle is intuitive and can
performances have been evaluated by using simulation. In
the simulation, data sets ranged from extreme small to large
apply to the most real situation (but not always).
have been considered. Weibull distribution with increasing, Suppose the observed failure time are , , … . , , . . . .
constant and deceasing failure rate have been chosen in the The likelihood function is defined as
simulation to ensure the simulation’s results concrete. ∏ (2)
Which can be rewritten as
Keywords – Maximum Likelihood method, Moment /
method Parameter Estimation, Simulation, Weibull ∏ (3)
distribution. The is a very small number in most cases that causes
inconvenience of computation, as shown in Table 1,
I. INTRODUCTION where the likelihood function value is very small.
Table 1. L vs ln L
Weibull distribution is most important distribution in
reliability engineering[1]. The major characteristics of
this distribution is that it can represent an increasing, 10 0.10 4.81E-130 297.7654 4.81E-130
constant and decreasing failure rate, as shown in Fig 1. 20 0.64 8.18E-102 232.7616 8.18E-102
This advantage endows this distribution a great 30 1.19 9,15E-103 234.9524 9,15E-103
importance in reliability[2-4].
40 1.73 1.07E-110 253.214 1.07E-110
4
… … … … …
Time vs Beta 2
Time vs Beta 3
3 Time vs Beta5
Time vs 0.8
One trick takes the logarithm of to transform the into
Failure Rate
2
an easily tractable number. Finding the maximum of is
equivalent to find the minimum ln , as ln is a
1
monotonic function of .
0 2 4 6 8 10
Time
Fig 1. Failure Rate of Weibull
ln ln ln / 1 ∑ ln ∑ Γ 1 (10)
(4)
For continuous function as (4), the simple method to find Since there are two parameters, the first two orders can be
the minimum is to take first derivative with respect to used to estimate parameters. In this paper, the method
each parameter, i.e. used to estimate the parameters are as
̅
(11)
∑ ln
Similar to the maximum likelihood estimate, numerical
∑ ln ∑ ln (5) method must be used to find the solution. Our
experiments shows the unique solution always exists so
Equal them to zero and simplify the expression. The (5) is the uniqueness of the solution can be guaranteed.
then rewritten as
The confidence interval of estimators can be derived from
/ moment also. According to large number theory, for a
∑ / ; series of i.i.d. data , if | | ∞, estimators of
∑ parameters
∑ ln 0; (6)
∑ , ,… , , , …, . (12)
Let and , , then
The depends on the . The can be obtained from
√ → 0, ′ (13)
Equation (6). The equation has not the closed-form
We omitted the proof of (13) in this paper. Variance of
expression. One has to resort numerical method, e.g. the
numerical Newton-Raphson method to find the can be approximated from (13). The remaining problem
now is that the explicit expression of
approximate . This is a weakness of MLE for Weibull
Distribution.Generally, MLE obtained from first , , …, is not available.The for (13)
derivatives like (6) does not guaranty the values obtained can be obtained Jacobian matrix. Let ,
is theminimal. Section VI will address this problem
specially. According to the Jacobian,
Confidence Interval Derivation / … / … /
.. … …
/ … / … /
The second derivative of (5) is , , …,
(14)
/ … / … /
ln .. … …
1 / … / /
…
ln
ln The first derivative of each implicit function can be
obtained from (10).The is the covariance of , . For
ln example, the . Using the
∑ ln / / (7) parameter value estimated from (11), a numerical matrix
of ′ can be obtained. In the matrix , the
diagonal corresponds the variance for each parameter as
The fisher information matrix [5] is
, (15)
803
Authorized licensed use limited to: UNIVERSIDAD POLITECNICA SALESIANA. Downloaded on September 30,2020 at 20:34:55 UTC from IEEE Xplore. Restrictions apply.
Proceedings of the 2015 IEEE IEEM
the to , in (9). 25
1/Beta
first derivative may not exist, or it may not be unique. As 10
Beta
6 8 10
are several solutions to the problem, the estimators will Fig 3. Cross of RHS and LHS of (18)
vary. Furthermore, the solution may not exist. The
estimator from numerical method may not have any The plot of RHS and LHS of (18) must cross to ensure the
sense. One has to ensure the solution exist and it is unique solution to (18) exists. Let us write it in terms of
before applying numerical method. mathematics, sufficient additional conditions for the
existence of solution to (18) are
In state of art, some authors claim that the MLE is not
unique. Some have proved it is unique, however their i. lim → lim →
proof is not complete and not mathematically
strict[7].This section proves the MLE exists and it is ii. lim → lim →
unique. Let us rewrite the Formula (6) as
∑
∑ ln It is straightforward to findlim → ∞ and
(18)
∑ lim 0 . It is obvious that the
→
Let write RHS of (18) as
lim → lim → . The condition i holds. For
∑
∑ ln (19) condition ii,
∑
The first derivative of lim → 0.
∑ ∑ ∑
(20)
∑ Let max , ,,… ,…. . The
Let ln and , the numerator of RHS ∑ ln ∑ ln
(20) is rewritten as
∑
∑ ∑ ∑ (21)
It is obvious
According to Cauchy-Schwarz inequality, lim 1;
0 (22) →
The equality holds when there exist constants and lim 0;
(not both of them are 0) for all the 1 →
(23) So
i.e. ∑ ln
Lim ln ln ln
→
ln (24) ∑
Simplify the (23) as ln /
Since / 1, the ∑ ln / 0
|ln | (25)
lim → 0 (26)
As for ,
The condition ii therefore holds. For two monotonic
The (24) holds only when and are both 0. This
curve crosses, the cross point is only one. It thus
contradicts to the condition of equality of (23). Therefore,
concludes that the MLE estimator exists and it is unique.
the (23) does not hold. The 0and 0. So
V. EQUIVARIANCE OF ESTIMATORS
0
Life data in engineering is common to change in it scale.
For example, the time to falure is perhaps in hours, or in
The in (19) is strictly monotonic. days, in weeks. The shape parameer that characterizes the
reliability should not change with the scale. The estimator
This monotonicity of RHS of (18) is not sufficient to for shape parameter should therefore unique regardless of
guaranty the existence of the solution. The plot of LHS of the scale. In statistics, it is named equivariance.
(18) is as the blue line
804
Authorized licensed use limited to: UNIVERSIDAD POLITECNICA SALESIANA. Downloaded on September 30,2020 at 20:34:55 UTC from IEEE Xplore. Restrictions apply.
Proceedings of the 2015 IEEE IEEM
Without losing the generality, we write atime to failure in (Maximum Likelihood Method). It is LCI in
the new scale as Table 2-4.
; 0 (27)
The is positive constant. The formula to estimate the The simulation results are shown in Table from 2 to 4.As
shape parameter in the new scale is the confidence interval is set for 95%. The error tolerated
∑ should be around 5%. However, for the extreme data size,
∑ ln (28) the error exceeds 5% significantly, around 15% as shown
∑
Substitute the (27) to (28). The constant is canceled and in Table 2. It is around 5% for middle data size as N=80
the (28) is equivalent to (18), i.e. they have the same as shown in Table 2. This means the confidence interval
solution. does not work for extreme small data size, only works for
The scale parameter is estimated from middle and large data sample size.
/
∑ / (29) Error of point estimation decreases, when the data size
As expected, the scale parameter scales with the scale of becomes bigger, as shown in the Fig 4. The error of CI
the life data. also decreases when data size bigger. However, the
The moment method use the following formula to CItend to be stable when the data size reaches around 100
estimate as shown in Fig 5. The CI length shows significant
difference between the two methods. The length from
∑
(30) likelihood method always outperforms moment method,
∑
especially when the data size is big as shown in Fig 6.
Substitute the to (30), it is readily to find
∑ ∑
. The solution to is thus equal. The scale Conclusively, for point estimation, for all the data size
∑ ∑
and Weibull models, the two methods has similar
parameter is estimated from
performance. For CI, the likelihood method outperforms
Γ 1 (31) the moment method, especially for the big data size.
Substitute the to (31), it is readily to find 0,25
Likelihood Method
0,15 Moment Method
0,10
Sample Size
represents the shape parameter equal to 1, greater than 1 Fig 4. MSE of against sample size
and less than 1. The sample size N is from the small
sample size to large sample size (N=5, N=25,
18
Moment
defined how close the estimators to the true parameters. Sample Size
∑
Percentage of length of M longer than L
95
(32) 90
85
70
Sample Size
(Moment method) is longer than that of L Fig 6. Percentage of CI Length of Moment longer than Likelihood
805
Authorized licensed use limited to: UNIVERSIDAD POLITECNICA SALESIANA. Downloaded on September 30,2020 at 20:34:55 UTC from IEEE Xplore. Restrictions apply.
Proceedings of the 2015 IEEE IEEM
806
Authorized licensed use limited to: UNIVERSIDAD POLITECNICA SALESIANA. Downloaded on September 30,2020 at 20:34:55 UTC from IEEE Xplore. Restrictions apply.