Thank you all for commenting and for your compliments! @Chet: I agree that setting up an ALT is a subject to errors of different kinds as it is generally impossible to bring the nature to the lab. For that reason, the lab tests should not solely be the means of validation. However, it is possible to measure (with limited precision, of course) the quality of the lab tests and their similarity to the real conditions. One way of setting the tests is therefore optimization of a measure of similarity (optimization for implications rather than for the induced loading). In that way we know what the limitations of the lab test are, and have that on hand when making conclusions about the test results. Your comment about the combination of the different failure modes was pretty inspirational. I know this is the generally recommended practice, reflecting on the often observed effect, that the individual failure modes (unlike all of them mixed together) better fit the usual distributions. I understand it is practical, relatively robust to insane extrapolations, the combined distribution (it is a weighted addition, not a multiplication) gives a good estimate of the “composite” reliability function. I would like to offer an alternative view of what is done by this: In the end we actually have a distribution function of many parameters. E.g. if we identify 3 failure modes, model them with 2 parameter distributions, then we end up with a distribution function of 11 parameters (3x3 parameters + 2 weights, the last weight is determined by the previous two). This is a rather complex function and we are likely to make some unnecessary assumptions. The more parameters we have, the more sensitive we are to assumptions. And it is just three failure modes. The quality of the parameter estimation is the lower, the more parameters we have (degrees of freedom problem). The other question is how we will justify if the observed failures are one failure mode or two or several. Sometimes it is apparent (e.g. brittle fraction versus abrasive wear), but sometimes it is not (fatigue failure in one weld, starting at just slightly different points). Technically, it is only a matter of detail we choose, that will make us to combine or separate them. For people who measure a reliability of a fleet of machines (e.g. rental shop or a mine) it would be “that bloody xy machine again”, because the xy machine is their component. While for the welding expert in the xy factory it would be two very different points of weld on one specific part, needing two different approaches to fix them - the component here is not just the one part that failed, but a particular portion of it. And they both would be looking at exactly the same failures. So fitting a reliability function of e.g. 5 parameters to all failures mixed together should be equally valid, and the quality of it will likely be better. I agree we need the reliability function to reflect on all of the failures and failure modes, but there is more than one way of obtaining it. As we depend on statistics (and statistics does not care about ships, planes, excavators, and trains, contrary, its virtue is abstractness), we always should try more than one way and see if we end up with the same conclusion. If we do, we know that the conclusion is not sensitive to the approach we choose, which is good, because we never know which way is the right one. The goal is what matters. BTW: There is quite a broad discussion on the video in the LinkedIn’s “Design for Reliability” group - have a look!
Now I want a ceiling that has relevant equations on it when I look up. Mine only has an occasional spider. The content was informative, but the presentation was exciting.
The testing of the redesigned replacement part was inevitably an Accelerated Life Test (ALT) which is subject to (at least) three kinds of error due to the acceleration: (1) artifacts of the test process introduced that are false failure modes, (2) real failure modes that are masked by the accelerated test, and (3) error in applying a Life vs. Stress relation with its all-important constants. Further as failures occur in the field or in the in-house Life Test, the failure modes should be separated and time-to-failure should be plotted separately for each failure mode, censoring (suspending) other units which fail for a different mode, or which don't fail at all. A distribution should then be fitted to the data for each failure mode. This would likely be a 2- or 3-parameter Weibull distribution which can emulate Exponential, Normal, Log Normal, and other distributions. The actual Reliability distributions, R(t), for each failure mode can be multiplied together to obtain a much better picture of reality. Simply taking the Mean, mixing together different failure modes, and not taking into account the times-to-failure is woefully inadequate.
Great presentation. An interesting topic and thought provoking question. Sadly, an all too often misunderstood measure of reliability. Perhaps I could point out some minor errors. Your flipchart suggests the units of MTTF is months but you describe years. Your Reliability scale (y axis) on your plots suggests (time) but this should be probability of success.
and what would you do, if you have to choose between 2 parts from different manufacturer: the 1st part has MTTBF=10 years, and another manufacturer has MTTBF =20 years. Which one you will prefere? Does it mean statistically, that the second one is more reliable? In the very begining you will always have early failures.
A little disagreement with the core points made by the lecture. The sum(TTFi)/N is not the reliability which used in statistics or engineering, it is just Mean Time To Failure.
Hi Ed! I was actually trying to speak English in the video and I hope I achieved at least a slight resemblance. My mother tongue is Czech. The glasses are “Parasite” brand of French origin, I bought them in Prague.
![NLE Choppa - Or What ft. 41 (Orchestra Mix) [Official Music Video]](http://i.ytimg.com/vi/biOFASEgSOM/mqdefault.jpg)
![Megan Thee Stallion - Roc Steady (feat. Flo Milli) [Official Video]](http://i.ytimg.com/vi/HozPPyqW0dY/mqdefault.jpg)
Thank you all for commenting and for your compliments!
@Chet: I agree that setting up an ALT is a subject to errors of different kinds as it is generally impossible to bring the nature to the lab. For that reason, the lab tests should not solely be the means of validation. However, it is possible to measure (with limited precision, of course) the quality of the lab tests and their similarity to the real conditions. One way of setting the tests is therefore optimization of a measure of similarity (optimization for implications rather than for the induced loading). In that way we know what the limitations of the lab test are, and have that on hand when making conclusions about the test results.
Your comment about the combination of the different failure modes was pretty inspirational. I know this is the generally recommended practice, reflecting on the often observed effect, that the individual failure modes (unlike all of them mixed together) better fit the usual distributions. I understand it is practical, relatively robust to insane extrapolations, the combined distribution (it is a weighted addition, not a multiplication) gives a good estimate of the “composite” reliability function.
I would like to offer an alternative view of what is done by this: In the end we actually have a distribution function of many parameters. E.g. if we identify 3 failure modes, model them with 2 parameter distributions, then we end up with a distribution function of 11 parameters (3x3 parameters + 2 weights, the last weight is determined by the previous two). This is a rather complex function and we are likely to make some unnecessary assumptions. The more parameters we have, the more sensitive we are to assumptions. And it is just three failure modes. The quality of the parameter estimation is the lower, the more parameters we have (degrees of freedom problem).
The other question is how we will justify if the observed failures are one failure mode or two or several. Sometimes it is apparent (e.g. brittle fraction versus abrasive wear), but sometimes it is not (fatigue failure in one weld, starting at just slightly different points). Technically, it is only a matter of detail we choose, that will make us to combine or separate them. For people who measure a reliability of a fleet of machines (e.g. rental shop or a mine) it would be “that bloody xy machine again”, because the xy machine is their component. While for the welding expert in the xy factory it would be two very different points of weld on one specific part, needing two different approaches to fix them - the component here is not just the one part that failed, but a particular portion of it. And they both would be looking at exactly the same failures. So fitting a reliability function of e.g. 5 parameters to all failures mixed together should be equally valid, and the quality of it will likely be better.
I agree we need the reliability function to reflect on all of the failures and failure modes, but there is more than one way of obtaining it. As we depend on statistics (and statistics does not care about ships, planes, excavators, and trains, contrary, its virtue is abstractness), we always should try more than one way and see if we end up with the same conclusion. If we do, we know that the conclusion is not sensitive to the approach we choose, which is good, because we never know which way is the right one. The goal is what matters.
BTW: There is quite a broad discussion on the video in the LinkedIn’s “Design for Reliability” group - have a look!
Excellent!! Very creative presentation. Excellent coverage of the material.
Now I want a ceiling that has relevant equations on it when I look up. Mine only has an occasional spider. The content was informative, but the presentation was exciting.
Very well done. I like the camera work, animation and very well explained use of statistics.
High Quality your video that pass a reliability. It's clear and very nice your job.
The testing of the redesigned replacement part was inevitably an Accelerated Life Test (ALT) which is subject to (at least) three kinds of error due to the acceleration: (1) artifacts of the test process introduced that are false failure modes, (2) real failure modes that are masked by the accelerated test, and (3) error in applying a Life vs. Stress relation with its all-important constants.
Further as failures occur in the field or in the in-house Life Test, the failure modes should be separated and time-to-failure should be plotted separately for each failure mode, censoring (suspending) other units which fail for a different mode, or which don't fail at all. A distribution should then be fitted to the data for each failure mode. This would likely be a 2- or 3-parameter Weibull distribution which can emulate Exponential, Normal, Log Normal, and other distributions.
The actual Reliability distributions, R(t), for each failure mode can be multiplied together to obtain a much better picture of reality. Simply taking the Mean, mixing together different failure modes, and not taking into account the times-to-failure is woefully inadequate.
This is art
Great presentation. An interesting topic and thought provoking question. Sadly, an all too often misunderstood measure of reliability. Perhaps I could point out some minor errors. Your flipchart suggests the units of MTTF is months but you describe years. Your Reliability scale (y axis) on your plots suggests (time) but this should be probability of success.
and what would you do, if you have to choose between 2 parts from different manufacturer: the 1st part has MTTBF=10 years, and another manufacturer has MTTBF =20 years. Which one you will prefere? Does it mean statistically, that the second one is more reliable? In the very begining you will always have early failures.
brilliant video
Nicely done!
A little disagreement with the core points made by the lecture. The sum(TTFi)/N is not the reliability which used in statistics or engineering, it is just Mean Time To Failure.
Well explained....:-)
It´s not better for reliability use MTBF?
MTTF is in years or months?
which language does he speak? (and where did he buy the glasses?)
Hi Ed! I was actually trying to speak English in the video and I hope I achieved at least a slight resemblance. My mother tongue is Czech.
The glasses are “Parasite” brand of French origin, I bought them in Prague.
the video is very interesting.
感谢创始人马夫罗季先生！#MMMGlobal #MMMExtra #Bitcoin #Bitcoins #BTC #BitcoinReview #BitcoinNews #BitcoinMining #BitcoinForum #BitcoinATM #BitcoinTrader #BitcoinTalk #WhyBitcoin
Is it 8 years or months?
needs to be submitted to Cannes
lol, so many likes on a video that is basically teaching engineers how to design for planned obsolescence.
Nerds acting Cool.