IEEE TRANSACTIONS ON RELIABILITY, VOL. 44, NO.

3, 1995 SEPTEMBER 477

Modeling Failure-Data by Mixture of

2 Weibull Distributions : A Graphical Approach

R. Jiang on their suitability for modeling a given failure data set. There
The University of Queensland, Brisbane are 2 cases: the shape parameters of the 2 sub-populations are,
D.N.P. Murthy
different, or
The University of Queensland, Brisbane
the same.
For both cases, parameter estimation depends on the relative
Key Wordr - Failure data, Modeling, Mixed-Weibull distribu- ratios of the scale parameters for the 2 sub-populations. These
tions, Parameter estimation issues are discussed in detail, and the approach is illustrated
through two examples involving real data.
Summary & Conclusions - The paper presents a graphical J&K characterize the WPP plots where the shape
approach to decide on the appropriatenessof a mixture of 2 Weibull parameters for the 2 sub-populations are different. Their results
distributions to model a given failure data set. It involves plotting contain several errors and hence differ from those presented
the data on Weibull paper. The plots are completely characteriz- here. We critically evaluate their paper and discuss the reasons
ed and the parameters estimated for various cases. Two examples why some of their results are incorrect.
involving real data illustrate the usefulness of the approach. The
results are compared with those of Jiang & Kececioglu (1992) and
Section 2 discusses the mixed Weibull distribution and its
the errors therein are pointed out. plotting on WPP. Sections 3 & 4 deal with WPP plots and
parameter estimation where the shape parameters for the 2 sub-
populations are different. Sections 5 & 6 give similar results
where the 2 shape parameters are the same. Section 7 discusses
1 . INTRODUCTION
the appropriateness of a mixture of 2 Weibull distributions to
model a given failure data set. Section 8 presents two examples
Graphical approaches have been used extensively in deter-
involving real data to illustrate the approach. Section 9 evaluates
mining whether a particular distribution is suited to modeling
critically the results of J&K and lists the errors in their paper.
a given set of failure data. Plotting paper for several distribu-
tions is available to carry this out [2]. The procedure is fairly
Notation
straightforward. One typically plots the failure data on the plot-
ting paper for the distribution selected. If the data fall roughly f(t), R ( t ) [pdf, Sfl for the mixed Weibull distribution
along a straight line, then the distribution selected is not judg- 1 index for sub-population; i = 1,2 unless otherwise
ed inappropriate for modeling the data. In this case, one can specified
estimate the distribution parameters from the plot. J(t), Ri(t) [pdf, Sfl of sub-population i
vi, Pi [scale, shape] parameter of Ri( t ) , all are positive.
Acronyms p, q mixing weight for sub-population [ 1, 21; p E (0,l),
p+q = 1
J&K Jiang & Kececioglu [l] P P1 = p2 (the special case)
WPP Weibull plotting paper.
k (172/01)p (41)
Of particular interest is WPP as the 2-parameter Weibull X ln(t)
distribution has been used to model a wide variety of failure Y ( d ln(-ln(R(t)))
distributions [3]. If the plotted data are scattered along a straight Y ( X ) M-W(eX)))
line, then the 2 parameters can be obtained from the slope & e curve: y ( x ) vs x
intercept. Often when the failure data are plotted on WPP, the ej approximations to C, j = 1,2,3
points are scattered along a smooth curve rather than a straight Li plot for sub-population i on WPP
line. Then the data cannot be modeled by a 2-parameter Weibull Z intersection of L1 & &
distribution. A common next step is to decide whether the data L, asymptote to e as x - -a
can be modeled by a 3-parameter Weibull distribution or a com- L, tangent to e at x = xI
bination of 2 Weibull distributions. Two formulations that have A,B,C inflection points of y(x) for p1 # p2
received some attention are: T inflection point of y ( x ) for P1 = p2 = p
xo. yo coordinates of point Q in the x-y plane, Q E {A, B,
the mixture model 11, and its references] c, 1, r>
the competing risk model [4]. y’, y ” [first, second] derivative of y(x).
This paper characterizes the plots of a mixture model for Other, standard notation is given in “Information for Readers
2 Weibull distributionson WPP as a first step towards deciding & Authors” at the rear of each issue.

0018-9529/95/$4.OO 0 1995 IEEE

478 IEEE TRANSACTIONS ON RELIABILITY, VOL. 44, NO. 3, 1995 SEPTEMBER

L, is:
y = y ( t ) = ln(-ln(R(t))). (3)

From (2) & (3),

Y(X) = Y: + B. (X - XI).

The appendix shows that:

y = y(x) = In(-ln(R(eX))). (4)
lim ( s l ( x ) ) = l/p,
WPP involves plotting y ( x ) vs x. Eq (4) defines the smooth ’-*lx,

curve C.
lim (s2(x)) = 0.
e is the straight line: x-

L1: y = y ( x ) = P1-(x- 1n(vl)), f o r p = l (5) Use (13) in (9):

lim ( y ’ ( x ) ) =
x- f m
PI.

When p2 > pl, the intersection, I , of L~ & b is important. An intuitive explanation for this asymptotic result is: x
corresponds to t -
0, and x -
SODcorresponds to t
---03
03;
&
since > P1, the tail for fi ( t ) dominates the tail for A ( t ) as
t - -
0 and as t 03. This implies that for very small and very
large t, fi ( t ) dominates A (t). As a result, for the asymptotic
(7) case, the mixture is effectively a single Weibull distribution and
hence C! is a straight line with slope P1.
The appendix also shows that:

3. WPP PLOTS: [& > 011

lim [exp(-P1 -x)~ln[p+q.e~p($,,~(x))]]
= 0. (16)
Notation x-+m

$ 1 . 2 ~ e x p ( ~ l . x ) / v P- exp(P2.x)/vP Using this in (8) yields the asymptotes to e as the two straight
q(x) ln(Ri(eX))+l,for i=1,2, and ‘null’ lines:
a P.01 + q.02.
This section characterizes e. Use (1) in (4)and simplify: y = 01 (x - ln(v1)) + In@), as - -03y
(17)

e: y ( x ) = p1.x L1: (3,as x - +m.

La in (17) is parallel to L1 but displaced vertically by ln(l/p).
+ ln[v;” - exp(-P1.x) .ln[p+q.exp($l, AX))II. (8)
The inflection points of y ( x ) are the values of x where y ” =
C passes through I in (7), and, 0. From (9),
JIANG/MURTHY: MODELING FAILURE-DATA BY A MIXTURE OF TWO WEIBULL DISTRIBUTIONS 479

C? has 3 inflection points ( A , B, C) ;the relationshipbetween

these points is indicated above. Figures 2 - 4 show the plots
for 3 ratios of q1/72;the computer results agree with the
analytic results.
The inflection points are obtained by solving (18) = 0. Unfor-
tunately, it is impossible to characterize these points analytically.
The y ’ ( x ) was plotted for a range of parameter values and
in each case, the shape is similar to that in figure 1, where y ’ ( x )
& s i ( x ) are plotted. In all the cases plotted, y ( x ) has 3 inflec-
tion points ( A , B, C ) .

t
f
2.0 --
-1 .e 0 .el 1 .e 2.8 3.0 4.0
X

0.0 -- Figure 2. Shape of e, q l e q2

-2.0 -1.0 0.0 x 1.0 2.0 3.0

[p=O.2, q 1 = 1 , p1=1.25; 12-2.5, &=2.75]

Figure 1. q(x) & y ‘ ( X ) VS x

71/72 = 1:

-2.4 -1.2 0 .e 1.2 2.4

I

71/72 > 1:

3.1 Well-Separated Mixture

J&K [13 define a 2-Weibull mixture as being well separated
when:
These results are used later in parameter estimation.
The characterization of WPP plots for a mixture of 2
Weibull distributions e is:
The asymptote for x -
-03 is La. we treat the 2 case differently. In either case, two points of
The asymptote for x -
+03 is L1. special interest along C? are t = q i ; ( x i , y i ) denote the coor-
C passes through I; the slope of C? at this point is @, dinates of these two points.
480 IEEE TRANSACTIONS ON RELIABILITY, VOL. 44, NO. 3, 1995 SEPTEMBER

.L Cl: y(x) = ln(-ln(p.Rl(eX) + 4)); (26)

Y ' = P'Pl'RI.In(Rl)/[(p.Rl + 4).1n(pmR1+ 4)1, (27)

RI R1( e X ) .

The appendix shows that the asymptotic slopes for C , are:

-8.0 t // The asymptotes are straight lines,

I t (17) as x - -00:

-4.s -3.0 -1.5

X
0 .e 1.5 3.8
y(x) = ln(-ln(q)), as x - +m. (29)

Thus C , approximates C for large negative values of x (small

t). From (29), C1 approaches a horizontal line for large x
(large t) . Figure 5 shows the plot of Cl for a set of parameter
values and confirms this.
(19)

3.0 --

0.8 .-

-3 .e -.
I

R ( v l ) = p.exp(-l) + q; WIZ)= q-exp(-l). (22)

-1.0 8.8 1 .€I 2 .a 3 .a
x
From (19),
Lp=0.4, ~ 1 = 1 @1=1.25;
, 72=5.O, &=2.75]
x1 = ln(ql), y1 = In([email protected](-1) + 4)); (23) Figure 5. Shape of e & e,, and Inflection-Point A
x2 = ln(q2), y2 = In(-ln(q.exp(-1))). (24)
At A, y"=O. The intersection of the horizontal line (29) with
Because p2 ,
P1, R2 - faster Rl ( t ) - 1, as - C defines a point which can be viewed as an approximation to
A. (Plots for a range of parameter values indicate that this is
0. Therefore, for 0 < I 4 q2,
a reasonable approximation.) This yields:
R2(t) 1.
yA ln(-ln(q)). (30)
Thus (1) can be approximated as,
3.1.2 71 S ~ 1 2

R(t) = p * R l ( t ) + 4. (25) exp(-(v1/q2)P2) = 0 , exp(-(v2/VY') = 1 . (3 1)

The plot of this approximation on WPP is: Use (31) in (20):

 ~

JIANGIMURTHY: MODELING FAILURE-DATA BY A MIXTURE OF TWO WEIBULL DISTRIBUTIONS 481

R(vd = p.exp(-l), R(0d = p + q.exp(-l). (32) At C,the slope of y ' is zero. The intersection of the horizontal
line (39) with the C defines a point which can be viewed as
The two special points on C are: an approximation to C. Plots for a range of parameter values
indicates that this approximation is reasonable. This yields,
x1 = ln(01), Y1 = ln(-ln(p.exp(-l))); (33)

x2 = 1n(q2), y2 = In(-ln(p + q.exp(-1))). (34)

Because P2 > PI , R2(t) -

0 faster than R,(t) 0 as t -
00 . Therefore, for q2 a t < 00, R 2 ( t ) = 0, and as a result,
- 4. PARAMETER ESTIMATION: [P2 > 011

(1) can be approximated as: If the data set is very large, then the WPP plot has suffi-
cient points on either side of I so that one can fit the asymp-
R ( t ) = p.R,(t); thus, (35) totes for x -
f03. However, when the mixture is well separated
and vl -4 q2 [ql % q2] then most of the data is to the left
[right] of I , and as a result, one can fit only the asymptote for
C2: y ( x ) = En(-ln(p-Rl(ex)), (36)
x - -00 [+a]. When q l = q2, the data are scattered about
equally on either side of I so that one can fit both asymptotes.
y' = P1/[1 + (ln(p)/ln(~1))1, (37)
This implies that the estimation procedure for the well-separated
R, E Rl(e"). case must be different from that for q1 = 92.
Once the data are plotted on the WPP [2], one can infer
whether,
The asymptotic slopes for C2 are:
01 = 02,
lim ( y ' ( x ) ) = 0, x -lim
x--m +m
(y'(x)) = P1. (38) 01 9 02,
01 Q 02,
The asymptotes are straight lines, based on the scatter of the data relative to I .

y ( x ) = ln(-ln(p)), as x - -03; (39) 4.1 71 = 02

(5) as x - +CO.
The points are scattered on either side of the intersection
point I.
This implies that e2is a close approximation to e for large
positive values of x (large t ) and that C2 approaches a horizon- Steps
tal line for large negative values of x (small ?). Figure 6 shows
C2 for a set of parameter values and confirms this. 1. Fit a smooth curve to approximate e.
2. Draw L1 & L,.
3. Obtain P1, q1 from the slope and intercept of L1.
4. Obtain p from the vertical separation between Ll & La.
5 . Obtain I from the intersection of C with L1.
6. Fit the L,. From its slope, compute P2 using (11).
7. Draw & (slope P2 passing through I ) and obtain q2
from the y-intercept of L2.

4.2 01 a 02

Most of the data points are to the left of I . Hence, the

estimation must be based on the properties of C to the left of I .

Steps
1. Fit a smooth curve to approximate e.
2. Locate A and use (30) to obtain q & p.
I 3. Obtain q1 & 0 2 from (23) & (24).
-3.8 -2.8 -1.8 B .e 1 .e 4. Draw the tangent at the left end of C to approximate
)I L,. Obtain PI from the slope of L,.
5 . Draw L 1 ,parallel to L, and separated by ln(l/p). The
[p=O.4, q l = l , 01=1.25; qp=O.2, /31=2.75] intersection of this with the fitted C gives I , and thus XI.
Figure 6. Shape of e & e2,and Inflection-Point C 6. Compute p2 from (7).
482 IEEE TRANSACTIONS ON RELIABILITY, VOL. 44, NO. 3. 1995 SEPTEMBER

4.3 171 p 72
Most of the data points are to the right of I. Hence, the
estimation must be based on the properties of C to the right of I.

Steps
1. Fit a smooth curve to approximate e.
2. Locate C. Use (40)to obtain p & q.
3. Obtain q1 & 712 from (33) & (34).
4. Draw the tangent at the right end of C to approximate
L1, ensuring that it passes through (ln(ql), 0). Obtain from
the slope of L1.
5. The intersection of L1 with the fitted C gives I, and
thus XI.
6. Compute p2 from (7). I
' x- X. '

We need vl # q2. Let q1 < 72.

Rl(t) = exp(-(t/vl)P) = exp(-k-(t/q2)P) = (R2(t))k.(42)

The asymptotes are straight lines,
Use (42) in (1):

Use (2) & (3) in (43): y(x) = 0.( x - ln(q2)) as x - +W. (52)

ln(p.R,k-l+q)
y ( x ) = pax + In (44) Eq (51) corresponds to La, and (52) corresponds to &. Figure
v! In (R2 1 8 shows the asymptotes for a set of parameter values, and L1
given by (5) with = 0 . The vertical separation between L1
Y'(X) = P*P*Sl(X)+ Q'P'SZ(X)
& is ln(k), and that between La & & is ln(p.k+q); they
increase with k.
= 0.(p.k.(R2(eX))k-1 + q).s2(x). (45)

The appendix shows that:

Figure 7 shows y ' ( x ) & s i ( ~ for

) a set of parameter values. /-
The plots are similar for other parameter values.
The appendix also shows that:

t
-1 . B 0.0 1.0 2.0 3.0 4.0

= ln(p.q;B + q-qib), (49)

X

v2=15,p=2.75]
b=0.4, v1=1.5,
lim ( y ( x ) - P.x) = -P.ln(qz). (50)
x-+m Figure 8. Shape of e
.IIANG/MURTHY: MODELING FAILURE-DATA BY A MIXTURE OF TWO WEIBULL DISTRIBUTIONS 483

It is not possible to obtain the inflection points analytical- 4. Obtain (p .k + q ) from the vertical separationbetween
ly. Figure 7 shows that there is only one inflection point. Plots 4 & La; this separation = ln(p-k +
q).
for a range of parameter values indicate the same. This 5. Draw a vertical line at x = 1n(q2);its intersection with
inflection-point is T with coordinates (xT,y T ). C gives y 2 . This yields (p.exp(-k) +q-exp(-1)) from (55).
6. From the results of steps 4 & 5 , obtain p & k. From
5.1 Well-Separated Mixture these, q l is obtained using (41).
Following J&K, the mixture is well separated when k %
1
1. For the two points of special interest on C (corresponding X

to t = q l and t = q 2 ) , use (41) in (20): X

P X
X

R ( q , ) = p.exp(-1) + q.exp(-llk), (53a)

I
X
X

0.0 .’
R(772) = p-exp(-k) + q.exp(-1). (53b)

From (23),

x1 = ln(ql), y1 = In(-ln(p.exp(-1) + q.exp(-l/k))); (54) /’

xz = ln(q2), y2 = In(-ln(p-exp(-k) + q.exp(-1))).

When k %- 1, (53) can be approximated by (22), and (xi,yi)
(55)

-7.8.
ii
t
-1.2 0.0 1.2 2.4 3.6
X

and (1) is approximated by (25). Figure 9. Shape of C? & e3,and Inflection-Point T

e3denotes the plot of this approximation on WPP. The
shape of C3 is similar to e,. e3approximates e for large -x
6.2 k %- 1
(small t) and approaches a horizontal line (29) for large x (large
t). e3is plotted for a set of parameter values in figure 9 and
illustrates the statement above. At T, the slope of y’ is zero. Steps
The intersection of the horizontal line (29) with C defines a point 1. Fit a smooth curve to approximate e.
which can be viewed as an approximation to T. (Plots for a range 2. Locate T. Obtain q (and hencep) from (57).
of parameter values indicate that this is a reasonable approx- 3. Obtain q1 & q2 from (23) & (24).
imation.) This yields, 4. The data indicate whether a tangent line (an approxima-
tion to the asymptote) can be fitted to the right or the left tail
of the curve. If the tangent line is fitted to the right tail, one
needs to ensure that it passes through (1n(q2), 0). No such
restriction is required if the tangent line is fitted to the left tail.
6. PARAMETER ESTIMATION: [pi = 02 = 01 The slope of tangent line gives 0.

For k = 1, the data are scattered on either side of T and

as such one can fit both asymptotes. When k%- 1, the data are 7. MODELING FAILURE-DATA SET
scattered either mainly on one side or on both sides of T. As
a result, the parameter estimation procedures for the two cases 7.1 Steps
( k = 1. k%-1) are different. The procedure for k% 1 can handle
1. Plot the data on WPP.
the situations where the data are scattered mainly on one side,
2. Draw a smooth curve, C, to fit the plotted data. Check
or on both sides.
the shape of C by steps 3 - 5, and use only the step that best
fits the situation.
6.1 k = 1 e
3. CASE: is similar to the curves in figures 2 - 4.
Steps
Model the data set by a mixture of 2 Weibull distributions
1. Fit a smooth curve to approximate C. with PI # 02.The plot also indicates whether,
2. Draw the asymptotic lines & La. %r
3. Obtain 0 and q2 from the slope and intercept of b. q1/q2 = 1, or
484 IEEE TRANSACHONS O N RELIABILITY, VOL. 44. NO. 3. 1995 SEPTEMBER

data arranged in increasing order along with the transformed

values used for plotting. The plot accounts for the censored data;
for further details regarding censored data, see [2, 71. Figure
10 is a plot on WPP. A smooth fit to approximate C indicates
Use the appropriate approach in section 4 to obtain the parameter that the data can be modeled by a 2-Weibull mixture with p1
estimates. You are then finished. < p2 and q 1 % q2 since most of the data are to the right of I.
4. CASE: C is similar to the curve in figure 8.
Model the data set by a mixture of 2 Weibull distributions
with common shape parameter 0.Use the approach in section TABLE 1
6 to obtain the parameter estimates. You are then finished. Data for Example 1
5. OTHERWISE: C cannot be modeled by a mixture of
Median
2 Weibull distributions. Look at formulations involving other Row rl Type' Position Rank R(rl) ln(t,) In(-lnR)
distributions to model the failure data set. You are then finished.
1 478 0 1.oooO 0.013889 0.9861 11 6.1696 -4.26968
7.2 Discussion 2 484 1 t
6.1821
3 583 0 2.0204 0.034135 0.965865 6.3682 -3.36012
The decision in steps 3 & 4 of section 7.1 is somewhat 4 626 1 * * t
6.4394
subjective as it is based on visual evaluation. As a result, one 5 753 0 3.0625 0.054812 0.945189 6.6241 -2.87580
can, 6 753 0 4.1046 0.075488 0.924512 0.6241 -2.54479
7 801 0 5.1467 0.096165 0.903835 6.6859 -2.29156
reject a mixture of Weibull distributions to model a given data 8 834 0 6.1889 0.116843 0.883157 6.7262 -2.08544
set when the data set did come from such a model, 9 850 1 * * * 6.7452 1

accept a mixture of Weibull distributions as an appropriate 10 944 0 7.2558 0.138012 0.86 1988 6.8401 -1.90708
model when the data came from a different model. 4 11 959 0 8.2117 0.159181 0.840189 6.8554 -1.75228
12 1071 1 t * 6.9763
This leads to the problem of incorrect selection of the model; 13 1318 1 t * 7.1839 *
the probability of this occurring is high when the data set is 14 1377 0 9.4458 0.181464 0.8 18536 7.2277 - I .60825
15 1472 1 * * 7.2944 *
small. In such cases, other statistical approaches also encounter
16 1534 0 10.6001 0.204367 0.795633 7.3356 - 1.47571
the same problem. This paper does not address this difficult 17 1579 1 7.3645 *
problem. Rather, it allows the data to decide whether it is to 18 1610 1 1 * 7.3840
be modeled by a mixture of 2 Weibull distributions based on 19 1729 1 * 7.4553
20 1792 1 I * 7.4911 *
a visual inspection of the plots on WPP. * *
This approach is consistent with data-dependent modeling, 21 1847 1 t
7.5213
22 2400 0 11.9468 0.231087 0.768913 7.7832 - 1.33645
where one takes a pluralistic view in the sense that a given data 23 2550 1 * 7.8438
set can be modeled adequately by more than one model for- 24 2568 1 * * * 7.8509
mulation. Jiang & Murthy [5] discuss this issue in the context 25 2639 o 13.3932 0.249786 0.740214 7.8782 -1.20126
of comparing model formulations each involving 2 Weibull 26 2944 0 14.8396 0.288484 0.711516 7.9875 -1.07776
27 2981 0 16.2960 0.317183 0.682817 8.oooO -0.96357
distributions.
28 3392 0 17.7324 0.345881 0.654119 8.1292 -0.85692
The graphical is a point in The 29 3392 0 19,1788 0.374579 0.625421 8,1292 -0.75645
plot yields some insight for model selection and provides in- 30 3791 1 * I
8.2404 t

itial estimates. The graphical method does not provide any s- 31 3904 o 20.6941 0.404646 0.595355 8.2698 -0.65663
confidence limits for the estimates. More sophisticated statistical 32 4443 1 * I
* 8.3991 *
need to be used to the model and to obtain 33 4829 0 22.2891 0.436292 0.563708 8.4824 -0.55649
34 5328 0 23.8841 0.467838 0.532062 8.5807 -0.46046
more accurate parameter-estimates and s-confidence limits. 35 5562 25,4791 o,499585 o,500415 8,6237 -o,36771
We are studying extensive simulations to evaluate the ef- 36 5900 1 * * * 8.6827 t

ficacy of the graphical approach for various sizes of data sets. 37 6122 o 27.1805 0.533343 0.466657 8.7196 -0.27160
38 6226 1 * * * 8.7365
39 6331 0 29.0128 0.569698 0.430302 8.7532 -0.17047
40 6531 0 30.8451 0.606054 0.393946 8.7843 -0.07092
8. EXAMPLES 41 6711 1 ' 8.8115 t

42 6835 1 * I
* 8.8298 t

Notation 43 6947 1 t
* 8.8461 *
44 7878 1 * * 8.9718 *
implies an estimate. *
45 7884 1 t
* 8.9726 1

46 10263 1 * 9.2363
8.1 Example 1 47 11019 0 34.8761 0.686034 0.313966 9.3074 0.14710
The data are actual failures of the throttle for 25 pre- 48 12986 0 38.9071 0.766014 0.233986 9.4716 0.37328
* 9.4806 1

production prototype general-purpose load-carrying vehicles 49

50 l3Io3
23245 1 * * * 10.0538 1

from [6: exercise 10.191, The data comprise both failure and
right-censored observatgns. Here t iS the distance (h) before #o: Failure I : Censored
) ~ 1~ ; Data
failure (or censoring) as opposed to time. Table 1 displays the *Implies nor upplicubk
JIANG/MURTHY: MODELING FAILURE-DATA BY A MIXTURE OF TWO WEIBULL DISTRIBUTIONS

a .e

Y
a .8
Y

-2 .a

-2.5

-4 .a

-5.8

6.38 7.00 7.78 8.40 9.18

X
4.98 5.25 J.68 5.95 6.30
X
Figure 10. WPP Plot for Example 1
Figure 11. WPP Plot for Example 2

Use the method of section 4.3:

0 = 0.88, (4 = 0.12);
Ref [8] proposes a model with PI # &; its estimates are:
p^, = 1.3, = 9050 (km);
p = 0.5;

& = 5.0, dl = 301;

These are crude estimates and one would need to use more refin-
ed methods (eg, maximum likelihood) to obtain more accurate
estimates.
The advantage of our model is that it has one less parameter
8.2 Example 2 and the data do indicate that the 2 shape parameters are roughly
The data are actual time-of-death for irradiated mice given equal. The p can be estimated by the number of failures due
in [8: example 7.21. We confine our attention to deaths due to to each of the two causes: 0 = 22/60 = 0.367 (4 = 0.633).
thyamic lymphoma (22 data points) and reticulum cell sarcoma Our estimate for p is closer to this value than the estimate in [8].
(38 points). These data are given in table 2, and are plotted on
WPP in figure 11. The data are modeled adequately by a
2-Weibull mixture with common shape parameter 0. 9. RESULTS OF J&K

TABLE 2 J&K studied the plots on WPP when item failure is modeled
Data for Example 2 [8: p 1931 by a mixture of 2 Weibull distributions and 0, # Oz. They use:

Thymic lymphoma (22 data points) an analytic approach to characterize the asymptotic behavior
of the plots,
159 189 191 198 200 207 220 235 345 250
256 251 265 266 280 343 356 383 403 414 a computational approach to obtain the shapes for a range
428 432 of parameter values.
Reticulum cell sarcoma (38 data points) Based on this, they propose a classification scheme for the
317 318 399 495 525 536 549 552 554 557 shapes of the plots. They comment on the estimation of the
558 571 586 594 596 605 612 621 628 631 parameters of the model.
636 643 647 648 649 661 663 666 670 695 This section critically examines the J&K results and shows
697 700 705 712 713 738 748 753
~~
that:
some of their analytic results are incorrect,
Use the method of section 6.2: their inference based on computational studies are not correct,
some of their statements regarding parameter estimation are
0 = 0.3, 0 = 7.6; not valid.
186 IEEE TRANSACTIONS ON RELIABILITY. VOL. 44. NO. 3. 1995 SEPTEMBER

Four main points are treated in sections 9.1 - 9.4. plotted the curve for values of x to the left of I , then it would
9.1 Point 1 have been similar to [ 1: figure lb] . The shape of both these
J&K deal only with P1 # P2. This paper deals both cases, plots is similar to that in our figure 4, in the sense that they
and as indicated herein, the two cases need to be treated are close to & near I and approach LI as x - 03. The inter-
separately. val of x over which the curve flattens out (and is close to the
horizontal line y = ln(-ln(p)), shown more clearly in our
9.2 Point 2 figure 6) depends on q l / q z . The larger the ratio, the greater
is the flatness as shown in [ 1: figure la] relative to [ 1: figure
The main analytic results of J&K (proved in their appen-
lb].
dix) are:
[ I : figure IC]. This is similar to our figure 3 as q1 = q 2 .
- PI -
2A. y'(x) asx -00,
[ I : figures Id - lfl. These show the plots for values of x to
the left of I and correspond to q 1 < q 2 . Had J&K plotted
- P2 -
2B. y'(x) asx +00.
the curve for values of x to the right of I , then they would
have obtained shapes similar to that in our figure 2. The only
difference in [ l : figures Id - lfl is the interval over which
They prove 2A & 2B using a limiting argument which requires
q1 - -
0 and q2 m; this derivation is not valid and can lead
to wrong results. The results of our section 3 show that 2A is
the plots are relatively flat and close to the line y = In(-
ln(4)). The degree of flatness increases as q1/q2 decreases
as shown in [l: figures Id - lfl.
true and that 2B is false. Discussion. As argued in the 3 previous paragraphs, the shape
of the [1: plots] are similar for all combinations of P1/P2( #
9.3 Point 3 1) and q 1 / q 2 .The only thing that is different is the flatness
For p = 0.3, J&K compute C for a range of parameter of the plot near the inflection point. The inflection point is
values. Based on their extensive computation, they classify 6 to the right of I when q 1 / q 2> 1 and to the left when q 1 / q 2
shapes of C (labeled A - F) and show them in [ l : figures la < 1. The degree of flatness around the inflection point
- If]. Four comments on their plots are given in sections 9.3.1 depends on q1/q2. In light of this, the J&K classification
- 9.3.4. scheme has no basis.

9.3.1 Comment 3.1 9.4 Point 4

In [l: figures la & lb], the asymptotic tangent lines Q, ( t ) J&K discuss the problem of parameter estimation using
& Q2( t ) are marked incorrectly and need to be interchanged. their plots. They rightfully point out that the method of Kao
[5] and Cran [6] are incorrect and state that the method of
9.3.2 Comment 3.2 Jensen-Petersen [7] is not only correct but has a solid theoretical
basis. The method of Jensen-Petersen applies to q2 % q 1 and
In [l: figures l a - IC], y(x) ( e )approaches line Q , ( t ) is similar to that in section 4.2. It is not valid for q1 %- q2 or
as x -CO. Since the slope of Q, is PI, this disagrees with their
ql = q2. Section 4 explains why these cases need to be treated

-
analytic result which requires that C must have slope P2 as x differently.
00.

9.3.3 Comment 3.3

ACKNOWLEDGMENT
In [l: figures Id - lfl, y(x) seems to approach line Q2 ( t ) .
Since the slope of Q2 is 0 2 , this seems to imply that the simula- We are pleased to thank the Associate Editor, Dr. Kvam,
tion results agree with the analytic result as x - CO. This is and two referees for their constructive comments on an earlier
not true as indicated in our section 3. The reason for this false version of this paper.
impression is that their numerical computations were stopped
too soon. Plotting y(x) for x % 2 would lead to a different
conclusion. APPENDIX

9.3.4 Comment 3.4 We derive some of the results in the text. For notational
Our sections 3 & 5 establish that there are only two distinct convenience, we suppress the arguments and write Ri(e")as
shapes for C over the interval -00 < x < +m; the shape of simply Ri,etc.
C depends on whether PI = P2 or not. J&K claim 6 different
shapes; we comment on their plots [ l : figures l a - lfl and show A. 1 Preliminary Results
why this is not correct.
We state these without proof as they are easily obtained
[l: figures la & lb]. These are for q1 > q2. [l: figure la] using the L'Hospital rule. Interested readers can obtain a copy
shows the plot for values of x to the right of I; if they had of the proofs by writing to the second author.
JIANGIMURTHY: MODELING FAILURE-DATA BY A MIXTURE OF TWO WEIBULL DISTRIBUTIONS 487

Lemma 1 is for both cases, lemmas 2 - 6 are for the general

case P1 < &, and lemmas 7 - 9 are for P1 = p2 = p.

Lemma 1. lim (Rl.ln(R1)) = 0. use lemma 3 and the limit of sl:

X-+w

Lemma 2. lim (R2/R1)

x--m
= 1; lim ( R 2 / R 1 )= 0.
x-+w lim (s2) = lim
x- k m x-*-
( R2”n(R2)
R.ln(R)
) = O.l/p = 0. Q.E.D.

Lemma 3. lim ([R2.1n(R2)]I[Rl.ln(R1)])

= 0.
x--fw
A.3 Derivation of (15) & (16)

x-*w
+
Lemma 4. lim ([ln(R1) l]/[ln(R) + 11) = 1 . ($1,2(x) is defined in section 3)

Lemma 5 . lim (ln(R)/ln(R1)) = p ;

x--w

lim (ln(R)/ln(Rl)) = 1.
x- +a

Lemma 6 . X--W
lim (ln(Rl)/ln(p.R1+q)) = l / p .

Lemma 7. x--w
lim ( R 2 / R ) = 1; , ? y m ( R 2 / ~ )= 114.

Lemma 8. lim (1n(R)lln(R2)) = p . k + q ; Use lemma 5:

x--w

lim (ln(R)lln(R2)) = 1.
x- w

Q.E.D.
Lemma 9. x--w
lim (In(p.R$-’+g)/ln(R2)) = p . (k-1);

A.4 Derivation of (28)

lim (ln(p.R2k-’+q)/ln(R2)) = 0.
x-+m

Use lemma 6 in (27):

A.2 Derivation of (13)
lim ( y ‘ ( x ) ) = lim
Use L’Hospital rule: x--w x--00

= p*p1.1.1/p = 01.
lim (sl) =
x-fW
liyW
Use lemma 1 in (27):

lim (y’(x)) = lim

x- m x- m

= p.PI-O-l/[q-ln(q)]= 0. Q.E.D.

A.5 Derivation of (46) & (47)

Use lemmas 7 & 8:
Use lemmas 3 & 4:

Since,

Similarly,
488 IEEE TRANSACTIONS ON RELIABILITY, VOL. 44. NO. 3. 1995 SEPTE,MBER

3, 1995, pp 187-198.
lim (sl) = lim [6] A.D.S. Carter, Mechanical Reliability (2nd ed), 1986, pp 303-308,
x--00 x--m
455-460; John Wiley & Sons.
[7] K.C. Kapur, L.R. Lamberson, Reliability in Engineering Design, 1977,
pp 314-317; John Wiley & Sons.
[8] R.C. Elandt-Johnson, N.L. Johnson, Survival Model and Data Analysis,
1980, pp 192-196; John Wiley & Sons.

AUTHORS

= k . 0 . ( l / q ) = 0. Q. E.D. R. Jiang; Technology Management Centre; The Univ. of Queensland; Q 4072

AUSTRALIA.
R. Jiang received his BE & ME from Wuhan University of Water
A.6 Derivation of (49) & (50) Transportation Engineering, and is on the faculty of Changsha Communica-
tions Institute - both in P.R. China. He is a PhD student at the University
Use lemma 9 in (44): of Queensland, and has over 10 publications in marine machinery, automobile
engineering, and reliability.

Dr.D.N.P. Murthy; Technology Management Centre; The Univ. of Queensland;

Q 4072, AUSTRALIA.
D.N.P. Murthy obtained BE (Hons) from Jabalpur University and ME
lim ( y ( x ) - 0 - x ) = In[(1
x-+m
+ 0)/17!] = -/3.1n(q2).Q.E.D. (Hons) from the Indian Institute of Science, and MS & PhD from Harvard
University. He is Director of the Technology Management Centre at The Univer-
sity of Queensland. He has held visiting appointments at several American univer-
sities. His research interests include reliability, maintenance, quality manage-
ment, warranties, service contracts, and topics in operations management. He
REFERENCES has (co)authored over 100 research papers in international journals and over
70 conference papers. He is coauthor of two books: Marhemarical Modelling
[l] S. Jiang, D. Kececioglu, “Graphical representation of two mixed-Weibull (1990) and Warranty Cos?Analysis (1993). He is co-editing a handbook on
distributions”, IEEE Trans. Reliability, vol 41, 1992 Jun, pp 241-247. warranties to be published by Marcel Dekker in 1995. He is a member of several
[2] W. Nelson, Applied Life Data Analysis, 1982; John Wiley & Sons. professional societies, is on the editorial boards of three international journals,
[3] W. Weibull, “A statistical distribution function of wide applicability”, and has served on the advisory boards of several international conferences.
J. Applied Mechanics, vol 18, num 3, 1951, pp 293-297.
[4] J.D. Kalbfleisch, R.L. Prentice, ”he Sratisrical Analysis ofFailure Time Manuscript received 1994 November 4.
Dura, 1980; John Wiley & Sons.
[5] R. Jiang, D.N.P. Murthy, “Reliability modelling involving two Weibull
distributions”, Reliability Engineering and System Safety, vol 47, num IEEE Log Number 94-13254 4TRb

I’KOC,’EEUINGS FREE PROCEEDINGS FREE PROCEEDINGS FREE PROCEEDINGS FREE PROCEEDINGS FREE I’IIOCELIlIh’(;.\

Your Reliability Society gives each member a copy of the Proceedings for:
AR&MS (Annual Reliability and Maintainability Symposium),and/or
IRPS (International Reliability Physics Symposium),
depending on the type of membership.

IVwhen there are surplus copies, the Society tries to make them available to:
Instructors in Reliability. (Sometimes we can supply a copy for every member of a reliability class.)
Technical Libraries.
At this time, there are no such surplus copies. 4TRb