Reliability analysis … now what?
Speaker: Chris Jackson
Let’s say that you have some ‘reliability guys’ do some reliability analysis on reliability data … and they give you some numbers. Or a plot. Or a curve. Perhaps you are the reliability guy … and you do the analysis – and then what?
How do you use that probability plot to help you make a decision? How do you use confidence bounds, ‘p’ values, standard deviations, and so on to help you make a decision? The three most important things for reliability analysis are ‘the decision,’ ‘the decision’ and ‘the decision.’ So how do you convert reliability data analysis into useful information for that decision? … uncertainty and all? This webinar is for you!
This Accendo Reliability webinar originally broadcast on 25 August 2020.
The audio track is now an Accendo Reliability Webinar Series podcast episode. View the episode show notes to listen or subscribe to the podcast.
Please login with your site registration to view the video recording.
If you haven’t registered, it’s free and takes only a moment.
FOLLOW UP #1.
Hi All! Just had a couple of follow up questions that I thought I would answer collectively. The first one is an easy one! I talked in my webinar about a ‘2 year warranty period.’ So the time period we were exclusively focused on in the webinar was 2 years. But … on the charts, I had the period indicated at 0.2. This is my fault! I regenerated some plots based on some previous work … where the unit of time was decades. So 0.2 decades becomes 2 years.
Hope this explains this! If I redo this webinar in the future, I will fix this (obviously confusing) little detail.
FOLLOW UP #2.
I got asked a question about the ‘expected warranty cost per unit.’ If you look at the video at around the 29 minute, 50 second mark, you will see that there is a histogram of warranty cost per unit that appears to be ranging from around $0 to $30.
But … didn’t we have a warranty cost of $ 115 per unit?
Sorry for the confusion! We indeed have a warranty cost of $ 115 per unit IF THAT UNIT FAILS. However, on average we will have somewhere between 0 and 25 % (ish) of our products failing in the warranty period. Not every one will fail.
So, ON AVERAGE, we will expect to pay $ 115 multiplied by the EXPECTED WARRANTY FAILURE PROBABILITY for every unit sold. So if we expect 10 % of our products to fail, and we sell 100 of them, we would expect 10 products to fail (total). This will create a (total) warranty cost of $115 multiplied by 10 (failed products). Which means that out of our 100 products, we will EXPECT to pay $ 1 150 in warranty costs.
Or $ 1 150 warranty costs per 100 sold products.
Or $ 11.50 warranty costs per 1 sold product.
Which is why the warranty cost PER PRODUCT has to take into consideration the warranty failure probability, meaning it can’t be $ 115.
Please let me know if there is anything else you would like to ask me!
Dear Chris,
I understand that Weibull is best to plot only for single failure mode or failure mechanism. However, warranty returns can happen due to different kinds of failure modes. So how to calculate warranty cost in that case ?
Do i need to have weibull for each failure mode and then calculate individual warranty cost and then club them together for same period of time.
Ankur,
The Weibull is quite useful for single failure modes/mechanisms … but it is also really useful for single DOMINANT failure modes/mechanisms. As in, there is on failure mode/mechanism that drives most failures. So even if you have multiple known failure modes/mechanisms, the Weibull distribution could still be the best bet.
If there are multiple ‘dominant’ failure modes/mechanisms, you can create a likelihood function based on (for example) a series system of multiple failure modes/mechanisms.
I think you are asking if you do separate analyses based on each failure mode/mechanism … that is one way of doing it. You would then have to randomly add the failure probability samples of each analysis to come up with your whole system failure probability.
BUT … before you do that – confirm that you have these multiple failure modes/mechanisms. One thing that will suggest this is the case is where you create a Weibull plot and identify different regions of straight lines.
Let me know if you need anything more!