Page 370 - Failure Analysis Case Studies II

355
because there are long time intervals involved. In addition the test cylinders may have to be given
deliberately cut notches which may not accurately represent the real case.
The alternative procedure is to estimate growth parameters from cylinders removed from traffic
which have large defects in them such as those reported above. The problem with cylinders removed
from traffic is that there is only approximate knowledge about their service life, in particular
temperature and pressure history. For example in the two cylinders used in this paper, cylinder A
was 10.5 years old at failure and operated in the South of Australia, and cylinder B was 5.5 years
old at failure and operated in the tropics. It is not known how much of the time the cylinders were
at full pressure.
To obtain an estimate for the growth model it was assumed that cylinder A grew to the point of
leakage in 4.5 time intervals and that cylinder B took 5 time intervals to reach the critical size from
whence it failed during test. The time interval was set to one year which presumes that neither
cylinder was continuously filled during its service period.
Crack growth phenomena was assumed to be the same in both cylinders and that the operating
temperature was not a feature (a difference of 10°C in operating temperature makes for a difference
of a factor of about 4 from the Lewandowski data). While these assumptions may seem to involve
significant approximations, these approximations would have an effect of less than a factor of 10.
To put this in context, the existing estimates of crack growth rate are several orders of magnitude
incorrect for these two cylinders.
The equation predicted by this method has the following form:
log,,(da/dt) = 0.346KI - 0.027, (7)
that
where da/dt is in mm/h. The ratio of this result to the model given in Eqn (1) is 10(o.1'2K~+7.6),
is, a value exceeding IO8.

6. DISCUSSION

The above model cannot apply to the vast majority of cylinders in traffic or there would be far
more failures than have actually been experienced. The situation is that many cylinders have small
visually observable defects at the neck (for some manufacturing periods up to 10% of cylinders
have been questioned by inspectors). If the growth rate described above applied to all these cylinders
then leak and violent failure would be more prevalent. It is thus necessary to identify what is unusual
about the cylinders which fail or are detected with very large defects. This, in the view of the authors,
still has not been done.
The main proposal in the literature is that the lead level in the aluminium will explain the situation,
but this is not confirmed by the Australian cylinders which have always had low levels of lead
(below 10 ppm where measured in this work) and where recycled scrap has never been used. Fatigue
cycling is one issue which has been briefly studied. Stark and Ibrahim [8] show that 100 short cycles
increased crack growth rate by two orders of magnitude. Since all cylinders in traffic are subjected
to pressure cycling, it would be necessary to identify what type of cycle puts a cylinder at special
risk. It also has been identified that a very large grain size has an effect on crack growth [l], but this
is a characteristic of most aluminium cylinders. Incorrect filling, or filling with corrosive species, is
another possibility, but none of the cases studied have shown any indications that this may have
occurred.

7. CONCLUSIONS

In this paper a new approach was used for modelling sustained load cracking in aluminium gas
cylinders. The earlier models of crack growth were based on specimen testing which predicts
extremely low or zero crack growth rates when applied to actual cylinders. These models cannot
predict the failure of the cylinders which have actually occurred.
This paper presents a method of estimating constants in the crack growth equation by examining
actual failed cylinders and matching the theoretical model to the observations. The new equation
which is derived gives a crack growth velocity up to IO8 faster than earlier models.