RELIABILITY ENGINEERING

WEIBULL ANALYSIS OF LIFETIME DATA

Although, as explained in section 2, the Normal pdf can be used to represent wear-out failure,
the Exponential pdf purely random failure and the Hyper-exponential pdf running-in failure
(and statisticians have, in fact, conjured up many other such pdfs for these purposes), there is
one particular pdf, the Weibull pdf, which, because it can represent any of the three basic
types of failure, has been found to be particularly useful. In addition, it has two other practical
virtues, viz
(i) applicability via simple graphical techniques,
(ii) defining-parameters (the terms in the formula) which have an everyday
engineering significance.
The ideas underlying this pdf may be grasped from Weibull's own derivation, which was
neither mathematical nor statistical, but was based on a few practical considerations.

Weibull (a Swedish metallurgist working with the Bofors steel company) was involved in
analysing the results of load tests on many, nominally identical, test specimens of a particular
type of steel. Their Ultimate Tensile Strengths exhibited random variability, as they always
do. If F(x) was defined as the cumulative fraction which exhibited strengths less than a
particular load x (ie F(x) was the cumulative distribution function or cdf, the distribution of
the probability that a specimen would fail under the load x), then a plot of F(x) looked like the
one shown in Figure 1. None failed before some given load x0 (the guaranteed strength) and
a few hung on to quite large loads.

Firstly, Weibull conjectured that it might be possible to represent such a cdf fairly accurately
by the expression
F(x) = 1 – exp {– (x – x0)},

where(x – x0) would be some function, as yet undefined, of (x – x0), eg 3 (x – x0) or (x –

x0)2 or whatever, which itself increased as x increased. This would give a plot which started
at x0 and approached F(x) = 1 asymptotically, as required. However, (x – x0) would have to
be such that it gave the right rate of rise of the value of F(x), and would have to be
dimensionless (because it is an exponent).

1
Weibull found that the form

 x-x 
 (x) =  0 
  
Where  was a characteristic load (determining, along with x0, the scale of the loads
involved) and  a curve-shaping factor, gave him an expression for the cdf

 x-x 
F(x) =1 - exp ( 0 )  
  

which enabled him to correlate his test data very well. In addition, the expression had some
other very useful properties, as we shall see.

In the reliability problems that we are looking at here the stressing factor is not load but
running time t, since new or since last overhaul. The Weibull cdf for times to failure is
therefore written as
 t-t 
F(t) =1 - exp ( 0 )  
  
2
From this, some simple mathematics then leads to the appropriate expressions for the Weibull
pdf f(t), Reliability R(t), and Hazard Rate Z(t), ie –

 (t-t 0 )  1  t-t 0  
F(t) = exp ( ) ,
   

 t-t 
R(t) = exp ( 0 )   ,
  
And


Z(t) = (t  t0 )  1 .
 

Each of the terms in these expressions has a practical meaning and significance.
The threshold time-to-failure, or guaranteed life t0: In many cases of wear-out the first
failure does not appear until some significant running time t0 has elapsed. In the Weibull
expressions the time factor is therefore the time interval (t - t0).
The characteristic life,  When t - t0 = , R (t) = exp (-1) = 0.37, ie  is the interval
between t0 and the time at which it can be expected that 37 per cent of the items will have
survived (and hence 63 per cent will have failed).
The shape factor, . : Figure 2 shows how the Weibull pdf of time-to-failure changes as  is
changed (for clarity, on each plot t0 = 0 and 

*If  is significantly less than one, the pdf approximates to the hyper-exponential,
ie is characteristic of 'running-in' failure;
*If the pdf becomes the simple negative exponential, characteristic
of 'purely random' failure;
*As  rises above a value of about 2, the pdf converges ever more closely to the
Normal pdf, characteristic of 'wear-out' failure.

3
NB For the first two cases t0 must be zero, of course; for the wear-out case it may or may not
be. Also note from Figure 2 that  characterises the consistency of failure occurrence. The
larger its value the greater is the tendency for the failures to occur at about the same running
time.

Exercise
Get out your calculator and some ordinary graph paper. Assign actual values to t0 , 
and  and, using the expressions for R(t) and Z(t) given on Page 3, plot graphs of R(t)
and Z(t) vs t.
Firstly, try: t0 = 0,  and  = 0.5, 1.0, 2.0, and 4.0 respectively (Values of t from 0 to
4.0, say, in steps of 0.25).
Then try: t0 = 2,  and  = 3.0 (Values of t from 0 to 8.0, say, in steps of 0.25)]

Weibull probability paper

How do we test whether observed times-to-failure look as if they could be plausibly
represented by a Weibull cdf? In the language of statistics, whether they look as if they have
been sampled from such a distribution? And if they do, how do we determine the values of t0,
and which will give the distribution which best fits the data? One easy way is to use
Weibull probability graph paper. There are several versions of this; we shall use the one that
is most popular in the UK, marketed by the Chartwell technical graph paper company (Ref

4
No. 6572 in their list). A full size sheet is reproduced on page 7. On this, the y-axis variable
is the

cumulative fraction failed, F(t), expressed in per cent, and the x-axis variable is (t - t0), in
whatever are the appropriate units of time for the particular item studied (as explained in Unit
2, 'time' in this context is a measure of usage and might appropriately be 'number of
operational cycles'). The axial scales are so arranged that if a theoretical Weibull cdf were to
be plotted on the paper (ie using values of F (t) calculated from the expression given earlier)
they would lie on a perfectly straight line. The following example shows how the paper is
used.

Weibull analysis of a large and complete sample of times-to-failure

One hundred identical pumps have been run continuously and their times-to-failure noted.
A Weibull fit to the data is required.
1. The data is tabulated as in Columns 1 and 2 of Table 1
2. Successive addition of the figures in Column 2 leads to Column 3, the
cumulative percentages of pumps failed by the ends of each of the
class intervals of Column 1.
3. Three or four possible values, thought likely to span the actual value, are assigned to
t0 (the guaranteed life). The resulting values of t - t0 are tabulated in columns
4, 5 and 6. NB In each case t is the time of the end of the interval,
eg in Column 4, Row 3: t - t0 = 1300 - 800 = 500 hours.
4. On the Weibull probability paper, the Column 3 figures are plotted firstly against
those in Column 4, then against those in Column 5 and Column 6 respectively.
The result is shown in Figure 3, which also includes plots for t0 = 700 h
and t0 = 1100 h. The value of t0 which results in the straightest plot, in this case
900 hours is the one which gives a Weibull cdf which best represents the data.

5
5. The characteristic life, , is the value of t - t0 at which the line fitted to the straightest
plot reaches the 63% failed level, in this case 600 hours. (NB t - t0 = 600 h
corresponds to a total actual running time of t = 1500 h, remembering that
t0 = 900 h)
6. As shown, a perpendicular is dropped from the fixed 'Estimation point' (printed
just above the top left-hand corner of the diagram) to the straight line fit. The point
at which this perpendicular intersects the special  scale at the top of the graph gives
the value of  for the best-fit cdf (in this case approximately 3.5, clearly pointing to a
wear-out mode of failure)
7. Note that the perpendicular also intersects another scale, marked F This gives the
value of the cumulative percent failed, in this case 49.8%, at the mean life. From the
main straight-line plot we then read off that at this value of F (t), t - t0 = 540 hours.
So, since the estimated t0 is 900 hours, we deduce that the estimated mean life is
900 + 540 = 1440 hours.

6
Exercise
Plot the pump failure data of Table 1, Unit 2 (ie the cumulative, F (t), data) on the
Weibull paper provided and hence obtain values of t0, and for the distribution.
(You should find t0 = 200 hrs, hrs and  = 4.0, APPROXIMATELY – Remember,
this is a graphical technique!)

Small data samples, possibly incomplete, ie 'censored' – the method of Median Ranks
In practice, it is most often the case that only a handful of times-to-failure have been
recorded. Indeed, the items under examination might be large and expensive and only a few
might yet have been made. In addition, some of the items might still be running, not having

7
reached the failure point (ie be 'suspended') or may have been withdrawn from the trial (ie
'censored') because, in their case, the test conditions were accidentally altered.

In this situation the results of any analysis will necessarily be subject to greater statistical
uncertainty. A Weibull analysis might still be needed, however, on the grounds that an
approximate result at the end of a fortnight may be of more value than a precise one obtained
by waiting for another three months.

In the above case, a technique in which so-called 'Median (or 50%) Rank' estimations of the
F (t) values are plotted on Weibull probability paper can then produce meaningful estimates
of t0, and  It has the additional advantage that, using published tables of 5% Rank and
95% Rank estimations, confidence limits can be assigned to the estimated parameters. The
following example shows how the technique is applied.

Ten springs are being tested to failure. The situation to date is as shown in Table 2.
A Weibull fit to the data is required.
1. The failure points are ranked in ascending order ( Table 3, Column 2 ) and classified,
in Column 3, as failed 'f 'or suspended (or censored) 's' ( Table 3, Column 3)

8
2. For the first failed item the 'New Increment' is calculated from the formula

N+1-(Order Number of previous failed item)

New Increment=
N+1-(number of previous items)

Where N = total number of items in the sample (in this case, 10). Since this is the first
failed item, the previous Order Number is zero. Also, in this case, the number of
previous items is zero and therefore the calculated New Increment is 1 (Column 4).
Note that if the first failure had been preceded by some suspended items the New
Increment would have been greater than 1, eg if the first two items had been
suspended, the New Increment would have been –

10+1-0   1.22
(10+1-2)

9
3. The 'Order Number' of the first failed item is then obtained from the expression
Order Number = New Increment + previous Order Number
ie in this case, Order Number = 1 + 0 = 1 (Column 5).
4. This procedure is repeated for all the remaining failed items, in succession, ie
Second failed item: New Increment =
10+1-1  1
(10+1-1)
Order Number = 1 + 1 = 2

Third failed item: New Increment =

10+1-2   1.125
(10+1-3)
Order Numbe = 1.125 + 2 = 3.125
and so on......
NB After a suspended item, or group of suspended items, the value of the New
Increment remains constant, and therefore need not be revised, until the next
suspended item or group of such items.
[Mini-exercise: Check for yourself, from the data in this example, that this is
indeed the case]
5. Having completed Column 5, the corresponding 'Median Ranks' (Column 6) are
calculated from the formula
Order Number - 0.3
Median Rank =
N  0.4
eg for the fourth failed item
4.250  0.3 3.950
Median Rank =   0.380
10  0.4 10.4
6. The Median Ranks, expressed as percentages, are then plotted against cycles (or time)
to failure on Weibull graph paper, as in Figure 4 and values of ,  and  obtained (in
this case approximately 8200 cycles, 7400 cycles and 3.0 respectively), the same way
as in the previous example. In this case t0 = 0, this giving a good straight line plot,
but in the general case t0 would be established by trial plots exactly as previously.
7. Tables of Median (or 50%) Ranks (derived via the formula used above - called
'Benard's Estimate’) and also of ‘5% Ranks' and ‘95% Ranks', for various sample
sizes and Order Numbers, can be found in various textbooks [eg in Andrews and
Moss, and in O'Connor, see bibliography at the end of Unit 4] and are also
appended to this unit. It is a simple matter to plot these Ranks, as well as the Median
Ranks, against the observed cycles (or times) to failure, thus obtaining (see Figure 4)
the 90% Confidence Band for the data, ie the band within which it is 90% probable
that the plot would lie that would be obtained from a very large number of items.

10
In the absence of suspended items the procedure is, of course, much more straightforward, in
that the Order Numbers are simply 1, 2, 3, 4 ... N.

11
Exercise
Ten identical lifts were operated under identical conditions. For each, the total number
of operating cycles achieved before breakdown was as follows –

Lift Cycles to breakdown

1 19 104
   2 7 104
3 44  104
4* 32  104
5 16  104
6 17  104
7* 24  104
8* 6  104
9 4  104
10 68  104

*Lifts 4, 7 and 8 had not failed at the time of the survey

Using the Median Rank method, fit a Weibull pdf to the data, evaluating and  (cycles
to breakdown). Also, using the Rank Tables provided –
(a) fit a 90% confidence band
(b) hence estimate the number of cycles for which we could be 95% confident that the
lift Reliability (survival probability) would be at least 70%

(Answers: From the data alone t0 is clearly zero. From the plot, = –  = 32
– 36 × 104 cycles. (b) From the 95% Rank plot it can be seen that at t = 5 ×104 cycles,
there is 95% confidence that F(t) should be at most 30 %, ie R(t) should be at least
70%)]

12
13
14
15