Rolly12 had an interesting post regarding infant mortality rates which developed into a debate about the usefullness of Weibull analysis. I would like remain with that part of the discussion.

I am a supporter of using technologies, however, my stance on this issue is that background + experience + competance will produce very good results. I believe that training, education and assessment is essential for the hands-on people. Low grade skills are tolerated in maintenance and it's not helpful. It would be more fruitful to make sure that maintenance staff are highly trained and skillful that try to analyse the results of their performance.
Regards,
Joe Mc Cormack

If we know what we don't know, how about about engaging somebody else who can?

I first noticed the Weibull analysis term in an article written by HP Bloch on the need to have a training plan for new/graduate engineers.

I think Weibull will be there for those who want to use it for their own advantage.

Hmm..OK, I still believe that there is a long way to go in improving the quality of the hands on people before the analysis of failures comes into the picture. I'm not knocking Weibull Analysis, however, for those who have used it successfully, how many failures are required before the information is usable.Can Weibull analysis be used for a single failure, if not, what is required. I can understand statistics being used for a collection of similar equipment, especially if the equipment is at different locations. The people at one site would not be aware of the failure patterns at other sites. At one location though, I can't see where there is a difficulty with a good maintenance person being able to surmise what the failure modes are similar equipment. Am I being naive.
Regards,
Joe Mc Cormack

Hello Joe,

Having been part of the Weibull discussion on the discussion regards infant mortality; I would like to contribute to your thread also.
Having studied lots of equipment failure data back in the UK (a few years ago now) I wish I had been given access to Weibull analysis. During any failure data analysis back then we focused purely on the MTBF and providing we did some maintenance prior to the MTBF we should be okay. Quite often we weren’t. If we had analysed the data further using Weibull we would have realized that our failures were occurring on different parts of the bath tub and we had been implementing the wrong type of approach to defect elimination. To me it’s quite simple; if you don’t fully understand how your components are failing then how can you implement the right strategy to prevent future failures.

Here is an example:

Component A:
1st installation lasted 10 hours
2nd installation lasted 100 hours
3rd installation lasted 190 hours
Component B
1st installation lasted 99 hours
2nd installation lasted 100 hours
3rd installation lasted 101 hours

What is the MTBF of each component and would they require a different approach to maintenance?

I hope this provokes further discussion in your thread.

Cheers - Gary

Gary,
The answer you are looking for is 100 hours.. but 100 hours is most unlikely to be the MTBF. It could be anything. Anyway, to keep the thought process going, lets assume MTBF is 100 hours.
Let me know what you think the answer is to your question about whether or not these items need the same maintenance.
Steve

Gary,

Failure of component A is random in nature, Preventive or Time-Based Maintenance is not applicable other options to use here will be run to fail if consequences is low and minor chance of secondary failures, Predictive Maintenance if the component shows sign of a potential failure or modification if you dont want to get too good at repairing the failure.
It is also possible that modification of component A will change the pattern to a wear out mode, hence PM this time will be applicable


Component B is considered as age-related or there is a wear out pattern and the component survives this period so the best option here will be to perform a scheduled replacement on the 98th hr placing this on preventive maintenance.

MTBF is not recommended as a baseline for determining the frequency of replacement or repair since MTBF is just the use of averages

Regards,

Rolly Angeles

From my perspective, Weibull or life data analysis is a tool. How much you get out of it depends on how much is put into it. As the saying goes, "Garbage in, garbage out." Although, I understand that the amount of data available may have been someone else's decision, and there may be other limitations such as time and money. Depending on the data, there is only so much information that can be gathered from it. I agree that education, training and experience are very important.

You can do a Weibull analysis with a single failure, but you have to assume a value for Beta so that you can solve for Eta using maximum likelihood estimation. With a small sample size like this, the analysis can only be taken so far.

David

quote:

MTBF is not recommended as a baseline for determining the frequency of replacement or repair since MTBF is just the use of averages

Hello Rolly,

Thank you for your reply to my question regards the two component failures A and B. I liked your thought process and thought the answer you gave was reasonable. From your reply you say MTBF is not recommended as a baseline for determining the frequency of replacement or repair since MTBF is just the use of averages. Could I ask you how you determine the frequency of replacement or repair if you don’t use the MTBF?

Cheers – Gary

Hello Steve,

I would say the MTBF is 100 hours. However, I think it’s rude to answer a question with a question and would really like to see your approach first, plus, I think it’s only fair I give others on the forum the chance to answer the question I posed before I post what my solution would be.

Cheers - Gary

Hello Gary,

My reply is quite long and I would like to take time to think about what I'm going to say, I'll reply tomorrow.

My warm regards,

Rolly Angeles

Gary,

For A and B, do you expect both components to have a constant failure rate?

Best regards,
David

quote:

What is the MTBF of each component?

Gary, have you ever done any statistics regarding hypothesis testing? You have a small sample of three data points that you are using to determine an average for a large population. Yes... you can say the average of the ones that have failed is 100 days but you would be very very courageous to say that the average of a large sample is going to fail with the same average. If you have another fail at any time other than 100 days the average has changed.

quote:

Would they require a different approach to maintenance?

Sorry to appear rude... I guess my answer is "I dont know". There is not enough information to make an assessment. As Rolly points out, knowing the MTBF does not help in this case. It seems you have some way of answering the question that I dont thus my question in response.

quote:

Originally posted by GLT:
MTBF is not recommended as a baseline for determining the frequency of replacement or repair since MTBF is just the use of averages

This is taken from my training materials on Meaningful Measures of Equipment's Performance -Understandint OEE, MTBF, MTTF, MTTR, MTBA, Failure Rate . . . . .

Collecting failure data to calculate MTBF in order to determine the maintenance interval is wrong and should not be done since MTBF is a measure of reliability.

Failures fail predominantly into 3 categories, age related which make up to 20% of all failures and the bigger portion will be a combination or random and infant mortality failures which constitute to around 80% of overall equipment failures. And for Age-related failures it is not MTBF but rather the remaining useful life that is significant when attempting to determine the best replacement or maintenance tasks to avoid failures. Hence, Preventive Maintenance should be based on the useful life and not on the average life which is MTBF. Why ? If we use MTBF as a baseline to determine the frequency of replacement or repair on our equipment it means that you will encounter failures before your PM will be due and this is not good.

Could I ask you how you determine the frequency of replacement or repair if you don’t use the MTBF?

First we need to understand that failure comes in 3 patterns : Infant Mortality Failures, Random Failures and Age-related failures.

For age-related failures, (Beta is greater than 1) this is easy as long as we can determine when the majority of parts mostly fail and survive then replacing the part before it fails will be the best time for frequency to replace.

For Random Failures, what is the best frequency. Random failures adhere to no frequency as their timing of failure varies, and not all random failures will provide a potential failure that it is on the verge of failing. Determining the MTBF and using MTBF for random failures is not recommended since this is just using the average, this is where Preventive Maintenance is at its weakest since PM will not be applicable. Options we can use will be the following :

Run to fail : if consequences is low and chances f secondary failures are negligible

Predictive Maintenance - for random failures with potential failures, what is important will be to determine the P-F curve. We perform vibration analysis regularly every quarter, but once vibration starts to increase quarterly monitoring will be change to monthly and weekly until we decide finally to stop the equipment.

Modification or Redesign : Here we perform some failure analysis and Root Cause Failure Analysis by determining the Physical Roots, Human Roots and latent Roots of the problem. Although not all, but through modification or redesign it is possible to change the failure pattern from random to a wear out or age related pattern. For example, a bearing always fail prematurely, failure analysis indicate that the inner raceway shows sign of misalignment, hence, laser alignment was perform and part of a regular PM schedule. After addressing misalignment, bearing achieved it's normal life distribution.

I have attached a file so you can visualize random failures for case 1: Let the period be given in years, wherein the maximum life this ball bearing achieved is 9 years and a failure of 10 bearings (same specs and size) fail on this period, but you likewise have failures every year. Now for this case, we cannot dictate a frequency of replacement since if we perform replacement on the first year then this is a very expensive task, what if the bearing can reach 3, 4 or 5 years then we wasted a good life of the bearing. (Just like in most PM they replace parts that are still of good and working condition). If we perform our frequency of replacement on the 9th year then expect a lot of failures before the PM is due.

Now upon performing a redesign and modification on the ball bearing the pattern change, but still there are some premature or infant mortality failures (CASE 2) but 94 out of the 100 bearing fail between the 8th and 9th year, clearly we can decide to have the bearing be replace on the 7 or 7.5 years, but if you get the average which is the MTBF we need to consider the failures at the beginning.

I hope you understand my point.

My Warm Regards,

Rolly Angeles

This message has been edited. Last edited by: Rolly12,

Presentation4.ppt (81 Kb, 57 downloads)

Hi Gary,
Its been a week now since you posted your example. Rolly has provided his answer, David asked for more information and I have said I can't tell based on the three data points provided. Can you show us please how using Weibull you can solve the problem.

Also you asked Rolly the following:

quote:

Could I ask you how you determine the frequency of replacement or repair if you don’t use the MTBF?

Can I ask you to let the forum know how you use MTBF to determine the frequency?
Looking forward to a stimulating debate on this topic.
Either I am completely wrong or I totally misunderstand Weibull analysis - or I am right - Weibull in industrial maintenanace application has a very limited value.
Regards
Steve
www.pmoptimisation.com.au

Here is an example of the Weibull distribution applied to a real world case:

The other day, during a visit to a hospital whose medical equipment maintenance is being subcontracted, I was asked to advise when PM should be performed on two similar autoclaves (an apparatus in which steam under pressure effects sterilization) if preventive maintenance is performed by in house personnel, in order to bring their availability to a maximum. The two autoclaves work 24 hours a day and the part in question and deserving the most concern was one condensed water trap located at the bottom of each apparatus. The time to restore the part takes 24 hours in average (call the company in charge, trip to the hospital and the replacement). If the task is accomplished by personnel working in the premises, the mean time for the restoration was estimated as being 4 hours.

I gathered the dates of the last 6 replacements of each apparatus. Subtracted every two dates and got 10 TTF (in days) all together (63, 92, 41, 37, 88, 77, 54, 44, 49, 24). First of all, being the time intervals so short, I recommended the causes to get investigated and eventually eliminated. Then, using the median-rank method, I found the Weibull parameters being: eta = 2.09; beta = 55 days and gamma = 9 days, for a R2 = 0.9776.

Then, I found 27 days to be the adequate time interval providing the maximum availability possible (81,90%) for a CM/(CM + PM) ratio or F(TTF <= 27 days) = 0,09. The hospital abandoned the CM policy applied to the two autoclaves for a while, implemented this PM procedure and is now tracking availability on a day to day basis. The hospital intends to confirm the estimated results and in case so, progressively extend the in house PM policy to some other pieces of equipment.

Without the Weibull analysis I wouldn’t have any foundations on how to set a PM interval.

Regards,

Rui

This message has been edited. Last edited by: <Rui Assis>,

Rui,
Thanks for the contribution.
I have created a histogram from your data and attached it. First thing I hope the average punter will be able to recognise is that the data is more easily understood in this format. (spreadsheet attached)
I have a few questions if I may:
1. Do you know or did you find out what the mechanism (s) of failure were?
2. When you say, you only took the last six failure points - it suggests there were more. If there were more, why did you not consider these?
3. When you did the analysis, how did you calculate the time taken to do a scheduled replacement - was that four hours or did you consider the job could be done faster than that?
Regards
Steve
www.pmoptimisation.com.au

HospitalData.xls (15 Kb, 36 downloads) Autoclave Failure Histogram

Before I answer you Steve, allow me to post a few more lines on the issue raised by Gary.

With regard to the question posed by Gary, I would like to say that from a strict statistical perspective, as Steve correctly pointed out, data are just a few leading to poor accuracy. Once the data is treated by the software “Weibull ++7” from Reliasoft, one gets the following results:

Data: 10, 100, 190
Best of fit distribution = a two parameters exponential: Lambda = 0.0077; Gamma = -7.8963 for an R = -0.9840
Mean life = 121.4436
Confidence bounds (Fisher Matrix method for a confidence level of 90%): Upper result = 326.4242; Lower result = 42.1419. Therefore, one gets a very wide range of values from which an estimate of the real mean life could be picked up. Depending on the circumstances and the consequences, a value of 42 might be elected if you accept a risk of 10% to overestimate. On the other hand a B20, for instance, would be: 20.96 (confidence interval for a confidence level of 90%: 66.7052 >= TTF >= 3.2694). Which value would you choose? Once more, if one is to be conservative, I would choose 4 perhaps.

Data: 99, 100, 101
Best of fit distribution: a three parameters Weibull: Eta = 3.3907; Beta = 4.0388; Gamma = 96.3775 for an R = 1.0000
Mean life = 100.0053
Confidence bounds (Fisher Matrix method for a confidence level of 90%): Upper result = 101.3742; Lower result = 99.0115. This time, one gets a very narrow range of values from which an estimate of the real mean life could be picked up. Depending on the circumstances and the consequences, a value of 99 might be elected if you accept a risk of 10% to overestimate. On the other hand a B20, for instance, would be: 98.9725 (confidence interval for a confidence level of 90%: 100.4805 >= TTF >= 98.0187). Which value would you choose? Once more, if one is to be conservative, I would choose 98 perhaps.

In conclusion, despite the data are just a few, there is much to be said. It all depends on what course of action is to be taken. This way, a decision can be founded on some grounds and then proceed with experience as more and more data are gathered or knowledge is acquired on the subject under analysis.

Best regards,

Rui

This message has been edited. Last edited by: <Rui Assis>,

Hi Steve,

Some more data and, I believe, the histogram will show the typical pattern of a wear out failure. It is already on the way…

1. I was told some substance used to block the mechanism. It is a LRU and replaced by people from outside who don’t like to provide informations…(!). Anyway I recommended an investigation in order to find out the cause or causes leading to a failure in such a short period. The PM time interval will be changed accordingly in the near future if some progress is obtained of course;
2. Yes there were more failure points but they were incongruent for the analysis and I just didn’t consider them;
3. I considered what the engineer in charge told me as necessary for the accomplishment of the job. I really didn’t calculate or even estimate anything myself. It seems too long but you have to account for safety and sanitary procedures (check-lists) which require other people besides the maintenance people to be present and confirm the job as properly done.

Would you like to alert me towards other directions? I would very much appreciate any contributions. Thanks in advance.

Regards,

Rui

This message has been edited. Last edited by: <Rui Assis>,

Found this article when searching the internet:
http://www.barringer1.com/wdbase.htm
What does it mean? Does it mean Weibull analysis is applicable to the range of equipment listed in there?

The article says
"Weibull databases require considerable use of engineering judgment for meaningful application of the facts stored in database format. Do not apply the data blindly without understanding the specific situation, as this database is solely intended for educational purposes.
But guess what... people use this data to create reliability models that they think are accurate.
Steve
www.pmoptimisation.com.au

Why reliability bigshots recommend Weibull analysis?