Steve,

You said

quote:

if an inspection is 100% effective, then only one inspection is required inside PF.

Exactly; but how does one get inside PF? The start point, i.e. the P point is unpredictable in most cases. If we know the P point, the F point can often be estimated.
In order to get inside the PF interval, it can be proved that if you inspect at 50% or less of the PF interval you will always get inside the PF interval. that is the logic of the old wife's tale.

Hi Vee,
In my view, we will never precisely know either P or F and in fact they are likely to be variables with means and variance.
When we set condition based intervals, what we are trying to figure is an order of magnitude about the rate of deterioration; or in practical terms, the frequency of inspection such that the failure mode will not occur unexpectedly and also provide us with enough time to rectify the fault without too much disruption.
I know you know that --- but

quote:

In order to get inside the PF interval, it can be proved that if you inspect at 50% or less of the PF interval you will always get inside the PF interval. that is the logic of the old wife's tale

This does not sound right. Again I am not a mathematician.... If I have a normal distribution, and the mean is say 1 month. Two standard deviations cover 95% of the population. It would only need a standard deviation of one week for 2.5% of the sample to be less than two weeks.
Perhaps I am using the wrong approach in looking at this. Can you let me know how your maths works.
Thanks
Regards
Steve

Steve,

You said Quote
When we set condition based intervals, what we are trying to figure is an order of magnitude about the rate of deterioration Unquote

Again, you are completely correct. One way of knowing the rate of degradation is to know the P and F points. If you had continuous monitoring, it is possible to know both the P and F points. But we dont have this luxury in most situations. So what does one do is I think the main issue. By setting the right condition monitoring interval, you can 'catch' the failure some time after P, but still have enough time to catch another point some time before F. Pages 64-68 in my book give you details.

quote:

Perhaps I am using the wrong approach in looking at this. Can you let me know how your maths works.

The PF curve is a physical degradation model, NOT a statistical model. These are two different, though related domains. The PF does not need sophisticated statistical analysis, though some people do make it very complicated by doing so. The standard deviation or confidence limits approach does not add much value to the simple physical analysis relating to degradation parameters such as vibration levels or delta pressure. These are absolute measurements, not predictions, so there is certainty about them.

With failure patterns D,E and F, which are statistical hazard plots, we face a decision dilemma. What is the logic of doing a PM task on a given day when we know that the hazard rate is the same in 2, 20, 200 or even 2000 days later? In such cases the statistical approach gives way to the physical degradation model, namely the PF curve. If we can measure some point soon after the P point, we KNOW that degradation has started (even though the hazard rate is constant). If we can get a second point on the PF curve, we can be pretty sure where the F point is likely to be, so we can plan the repair in time.

I sent the book over a week ago, it should be with you by now. If not, I can try and track its movement.

What is the failure pattern for HVAC system?

Hi Vee,
On the subject of your fine book, I was away last week and the book arrived. I read the intro and skimmed a few chapters the first night back. I need some time to get right into it and I have to say I am looking forward to this.
Thanks
Steve

quote:

In order to get inside the PF interval, it can be proved that if you inspect at 50% or less of the PF interval you will always get inside the PF interval. that is the logic of the old wife's tale.Steve

Vee,
I am trying to understand how you come to this statement.
PF will have a distribution and therefore if I take the case that PF is a normal distribution, I am intrigued how you can prove that inspection at half the PF will mean that the rate of decay will never be faster than half PF.
I am taking some time to consider the other things you have mentioned regarding connecting hazard rate and degradation. I have a feeling they are mutually exclusive in practical terms.
I think I need to think more on that one before I respond - and read your text on the subject too.
Rgds
Steve

quote:

What is the failure pattern for HVAC system?

At system level it has to be a bath tub. Just like your car. If you break down your HVAC into components and failure modes - you will see different patterns.... just like if you look at your car at car level then take tyres (tires) and look at that pattern as a lone set. Tyres have three dominant failure modes - punctures, leaking valves and wear. If punctures and leaking valves are random you will see a pattern on tyres that has a random pattern with a slight wear out. Now if you take the valve as a separate component, you may see a bathtub again. The curves depend on the level.
At HVAC level they will aggregate to bath tub though.
Steve

How about at the HVAC fan belt and the air filter?

Josh,
I have lots of data about fan belts from various plants ... remarkably most fanbelt data analyses ends up showing the pattern is bathtub.
Infant mortality is rife... then alignment and pulley condition seem to reduce life. Eventually some belts wear out. In some of our data the mtbf is about half what the estimated wear life is.
On filters.... you would expect a wear out pattern. However in practice, process upsets tend to reduce life. Most filters go on condition for this reason - or sometimes duplex filters are installed so that change over does not stop production for long - if at all. There was some discussion and some data on a topic Failure Patterns. I posted on 08 April 2006 03:38 AM
Hope this helps.
If you are studying HVAC - do you have any failure data yourself?
Regards
Steve

Hello,

I am very interested to learn about failure data of HVAC, including the AC motors?

thanks

I have the work orders. Is it possible to see yours, Steve?

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

Hi all,

I finally reached the end of examinations of all my students.

I’ve just finished the writing of the document that I attach to this post in which I briefly comment the inspections issue (constant versus variable time intervals between inspections) that we have been discussing.

There is one more alternative to be considered which is “no inspections”.

Costs of each alternative depend on so many variables that it is practically impossible to anticipate in what circumstances one of the alternatives “variable time intervals”, “constant time intervals” or “no inspections” is actually the most cost effective. Therefore, I concluded that, when facing a real world case and economics is an important criterion to be taken into consideration, one should always take the time to equate the three alternatives.

Regards,

Rui

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

Inspections.pdf (56 Kb, 26 downloads)

I would like to add a few points to the discussion between Steve and Vee on the PF curve, if I may…

At times, having caught the failure sometime after point P and because you cannot be sure where exactly point F is going to be in time, you can assume a statistical behaviour for point F based on stored historical data or information from testing or you can use physics-of-failure models or covariate models (proportional hazards models) if any is available for the specific case (see Charles E. Ebeling, “Reliability and Maintainability Engineering”, McGraw-Hill, 1997).

In the stats case, you can simply use a Normal distribution in three steps:

 First talk to people with knowledge and experience on the specific failure mode (operators, supervisors, engineers,…) and ask them what, in their opinion, the pessimistic PF interval and the optimistic PF interval would possibly be. This way, you get two estimated PF curves: PF1 and PF2;
 Second, determine the two parameters of a Normal distribution: Mean = (PF1 + PF2)/2 and 2*Standard Deviation = (PF2 – PF1)/2. This distribution will cover 95.45% of all possible F values;
 Third, let this random variable interact with other random variables described by other probability distributions in a simulation model whose purpose can be, for instance, to estimate the cost of a particular strategy of inspection.

In the physics or covariate case, I count on colleagues who are experts in equipment working at high temperatures (steam boilers, steam and gas turbines, etc.) who give me their estimates founded on specific mathematical models for each component failure mode and failure site based on the stresses, material properties, geometry, environmental conditions and conditions of use. The failure times can then be ranked and the most dominant one provides the component time to failure estimate (or the remaining life).

Regards,

Rui

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

With regard to the third point of my last post, you can choose point F in such a way that it implies a X% chance (or allowable risk) that point F will actually appear sometime earlier. For instance, suppose PF1 = 500 hours and PF2 = 600 hours. Therefore: Mean = (500 + 600)/2 = 550 hours and SD = (600 – 500)/4 = 25 hours. If you now accept a 10% level of risk of overestimating point F, then, by using EXCEL, you get: NORMINV(0.1; 550; 25) = 517.96 or approximately 520 hours.

Hope it is clearer now.

Rui

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

Rui,
You and Ebeling are no doubt correct, but these steps are not easy to achieve in practice. It is hard enough to get an estimate of the P-F interval (by talking to operators, maintainers), it is close to impossible to get estimates of the F point itself, however dedicated and sincere the people involved. The P point is not recorded in CMMS, the F point is not known when the corrective action is carried out before the F point (Resnikoff again!), so the data does NOT exist anywhere.
The maths we use have to be pragmatic, appeal to people on the shop floor and close to real life situations. With the quality of raw data being what it it usually is, no amount of statistical polishing will improve its basic quality. Much better to get reasonable estimates of P-F interval in the first place. One of the real oroblems is the variability in the P-F interval itself, especially when it is in days or weeks.
The important question is really how to ensure we 'catch' the P point and still leave enough time to plan and schedule corrective work BEFORE we reach the F point. If the variability in P-F is so large with std. dev. close to the mean, I would opine that it not practical to define an inspection interval with confidence. In turn this means we cannot catch the P point. In a long career in maintenance, seeing, smelling, hearing and touching equipment, I have not found it too difficult to get P-F estimates, mainly by talking to people close to the equipment. By the same token, I have not come across anybody who knew when the P point would occur, and once the P point had been reached, when the F point would occur.

Hi Vee,

I agree it is not an easy task to get an estimate of P-F interval but it is not impossible and it may pay off even in the short run, if you have the skills and the right maths to assist you in the process of decision making. I already explained how I do it in my last post.

I use to work for an Institute where a group of engineers are specialised in process plants static equipment (coal and fuel fired boilers, steam and gas turbines, water pre-heaters, condensers, super-heated steam piping, and the like) in industries like refineries, pulp mills, coal and fuel fired electric power generation. They perform inspections (at times ruled by norms or set empirically – moving to scientifically!) and estimate when the effects of fatigue, corrosion, or any other mechanism of deterioration are to become critical (F point), after having noticed preliminary signs of a failure in progress (P point). They use the latest high tech technology available and the method chosen for each case is the result of a compromise (balance) between precision required and cost of resources necessary. There are some useful data from sources gathered since a few decades, which I use, when available and trustful, and after having been cross-checked with operators and maintainers. From time to time, a whole new problem appears and only engineering estimates founded on physics-of-failure models can be of help.

I understand there are circumstances in practice where no data is available and things become quite hard: No covariate or either physics-of-failure, or recommendations from norms, or even trustful data are available and data bases are too expensive and you feel lost and very unhappy!. In this case you have no other choice and trust on your experience and/or intuition. Fortunately, I rarely face this kind of situation!

Hope to have been sufficiently informative.

Regards,

Rui

Rui,
You misunderstood what I said.

quote:

not an easy task to get an estimate of P-F interval but it is not impossible

is NOT what I said. I DO believe we can get P-F estimates, especially by talking to the operators and maintainers. What I feel is impractical is to define the F point itself, not the P-F interval. If we catch the onset of failure, shortly after the P point, we can project the likely F point, knowing the P-F (estimated) interval.

Rui,
You say,
quote]Fortunately, I rarely face this kind of situation![/quote]
Good for you. I have NEVER seen a database with P-F intervals. I guess this is partly becaues the Operating context varies so widely from case to case, there can never be a standard P-F. It depends so much on how we operate and maintain equipment. You obviously have access to som high-tech information, but I find it hard to believe the universal nature of your data.

No Vee, I didn’t say the data is universal. The data that I referred to come from the same equipment operating since long or from similar equipment operating in similar conditions. As I said before, as soon as we notice the onset of a failure (P point) the F point is estimated taking into account any existing data (and stats comes into play) or, if the case is brand new, the pertinent physics of failure.

Regards.

Rui

Rui,

quote:

as soon as we notice the onset of a failure (P point)

Unless you do continuous on-line monitoring, how do you detect the P point? One objective of knowing the P-F interval is to decide how often we monitor, so we can 'catch' the P point or sometime shortly after. This will allow us to plan a 'predictive' repair before the F point.
I dont see how you 'notice' the P point unless you monitor continuously. In that case, all you need is the variability measure of the P_F interval. This matters if the variability is large, but is less critical if it is relatively samll. I agree that variability can be defined with the stats.
In most cases, we dont have the luxury of on-line monitoring. We still need the P-F interval to help decide how often to monitor, so that we can within reason catch the P point. This is the issue we were discussing earlier, not the case where we already KNOW the P point