Design for Testability

The Design for Testability
or the DFT is a concept that is best known in the fields like integrated
circuit design or computer programing. In this video, I would like to emphasize
its analogy in mechanical engineering and specifically in system reliability. The general concept of DFT is to make the design in
such way that it is capable of being efficiently tested. As usual, let’s see an example. A steering assembly of a bus, which was used
for a long time, was showing adequate reliability. During a reasonable period after the design
life, there was typically one component, a bearing that manifested wear, which
would potentially develop into a failure. The design team recognized this bearing to be
the weakest link, and since the reliability of it was perfectly OK, they considered it to be
a benchmark for the remaining components. They proposed a new, progressive, design,
with considerably better steering dynamics, but also a lower life in the drive
links due to thinner composition. The life of the changed components was
intended to match the life of the bearing. From the reliability point of view, the
components that form the system should generally be of
a balanced and comparable life. Intuitively, it does not make sense to have
a component that barely survives the expected duty, side by side another one, which exceeds the life of
pyramids and which is a dominant cost contributor. We often hear the analogy of a chain,
which can hold as much as its weakest link, which illustrates the idea quite blatantly. Nevertheless, this approach has
its limitations and drawbacks. First of all, we need to debunk one myth. When there is a chain, or, more precisely,
a “system of serial components”, the reliability of it is not the
reliability of its weakest link. The actual combined reliability is much smaller. The popular example fails to account
for the variation in the probable life. Suppose we have a chain where all links follow
a Weibull distribution with a shape of 2.5. The chain has one link with the scale of 1000
cycles and 99 links with the scale of 3000 cycles. The weakest link has only a third
of the life for every probability level. For example, the B10 life is 400 cycles for the
weakest link and 1200 cycles for all the rest. However, the B10 for the whole
chain will be only about 170 cycles, less than a half of the weakest link,
with the shape parameter of about 2.4. This is due to the fact that there is quite high
probability that there will be at least one link out of the stronger 99, which will have lower
life, than the one supposedly weak link. This will happen in over 85% of cases. This was quite an extreme example, not very
often we have a system with 100 components in serial setup, but it illustrates the point. This problem will be lower when
there are fewer components and where the “stronger” links will
be much stronger than the weak one. Due to this effect, the reliability of the
whole steering assembly lowered, but since there was not a prior
knowledge about the new drag links, it was not possible to estimate how much. The whole assembly would need to be tested. The preliminary analysis suggested that the whole
assembly does not have an apparent weak point and the failure mode that would develop
under the current testing method, would likely be caused by imprecisions in the method,
rather than an estimate of the real condition. It is quite usual to develop the test
methods in a conservative way, so that it is reasonably harsher
compared to the real conditions. On the other hand, the harsher tests may
produce unrealistic early failure modes, especially when the design is really balanced
or when it has a low overdesign margin. In this case, the design team backed up and
used an older design of the steering assembly, to save the cost of developing
a whole new reliability test. This was one of the cases, which demonstrate that the reliability engineering is mostly suitable
for small gradual improvements of the design, rather than a game changer that
contributes avant-garde ideas. Maybe the new design idea was a good one, but we wouldn’t know unless we
choose to spend the money in testing. I have a great respect for design engineers, especially those who take on the
burden of advancing the technology, and I feel the injustice when the reliability
testing sort of stays in the way of progress. Mostly we can and should keep up with the
progress by finding new ways of efficient testing, but sometimes there is just not the
business case that would justify the effort. There will always be a little fight
between the progress and the certainty, but after all, no risk, no fun. Here is a little summary of what one should
keep in mind when designing for reliability, to mitigate these disadvantages. Designing a weak link into the
system may have actual benefits. We know what to test, the testing
can be specific and less expensive. The maintenance people know what
to look for – if the weak link is still OK, than the rest of the system is also OK. The higher is the past knowledge about
the components to be used in the new design, the more targeted and more
efficient will be the testing. The popular chain analogy is only approximate, the
real chain is actually weaker than the weakest link. When you design with respect to the weakest link,
make sure that it really is much weaker than the rest and that the chain is not very long. The fewer components, the better. The smaller is the margin between the real
system reliability and the desired reliability, the more precise and more
expensive will the testing need to be. When designing a new product, it is good
not only to focus on the new features, but also to understand how the new
features will be reliability-tested. The Design for Testability can
generally be summarized to: Make sure that the design is capable
of being tested in an efficient way.


  • Sasuke Sarutobi

    Thank you for the video! It's fascinating to see how reliability engineering is applied to designing maintainability into a system, and also to see how Lusser's Law applies when using probability distributions.

    Out of interest, how would you look at a similar problem in the case that failure modes of the serial components are not independent? For example, if the failure of the weak component introduced stress on another component that increased the probability of the second component shearing. Aside from having to test the second component when detecting failure of the first, what sort of tests and design adjustments would you consider?

  • Shashank agrawal

    Nice video!!
    Very informative


Leave a Reply

Your email address will not be published. Required fields are marked *