How best to analyze the calibration results for pressure switches (hi-lo pilots) as attached?
What frequency should I recalibrate them? Their age is about a year ago except the last three are several months old.

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

Pressure_switches.xls (18 KB, 14 downloads)

The recalibration is a failure finding task.

Josh,

You said these pressure switches are practically one year old, but were they calibrated at the time of installation? If yes, for those which are 'out' it shows that the calibration frequency should be less than one year. However, again I think a risk based approach should be taken and what is safety or quality critical should get the most attention. Any input from the supplier?

How often you take readings from those pressure switches and take descisions using those readings?
How much will cost (in terms of co$t$, safety, quality)a wrong descision based on a wrong reading?

Are all those switches of the same model? 8/23 switches out of calibration? 33% of the switches failing? Maybe a look again at the design specs, are we sure these switches should be used for the application?

Of course, the pressure switches were calibrated before installation.

Specwise, they are ok.

Initial freq is one yearly.

I'm reluctant to do trial and error to come up with the freq, if possible.

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

Can we apply the failure finding formula (from Vee's book) here:

Test freq (T) = -ln (2A-1)/lambda ?

If availability A=97.5% for pressure protection purposes, how can I get the lambda?

Can MTBF (lambda) be the sum of operating times for the 23 pressure switches divided by no of failures which is 8 in this case?

Or do I have to take the value for lambda from other data sources such as OREDA?

When installing an instrument we check the following sources to determine calibration frequency:
* Mfg recommendations at manuals or other techinical reference.
* If no mfg recommendation, Do we have another instrument like this (same or similar model) installed at the plant? If yes, (after checking the calibration history) apply same calibration frequency used at the similar unit.
* If no other source to get a frequency determination we use a "rule of thumb" frequency based on instrument criticality:
* 1 year for non critical instruments
* 6 months for critical instruments

After several calibrations re-evaluate if the frequency can be increased or decreased. Instruments which every time they are calibrated are found to be within tolerance may have its calibration frequency decreased (6 months to 12 months). If every time an instrument is inspected for calibration is found out of tolerance then the frequency is increased (6 months to 3 months).

Eugene, yes we can set the frequency by the trial and error method (or age-exploration) as suggested by you above. However, isn't there a better way to calculate the frequency after just 1 set of calibration exercise? Will the calibration results vary very much for subsequent calibration exercises?

I do not think it is right to change the calibration frequency just by a bad result. I will check several calibration inspections and then act if the observed tendency proves the frequancy is too long.

My point is taking descisions based upon historical or statistical data. Can we judge the performance of an instrument calibration program just by the results obtained in one calibration order?

Note: by instrument calibration program I mean here the calibration instructions/method & frecuency.

One has to decide how much data one need to make a decision. If after a instrument fails the calibration inspection and we shorten the calibration interval, do we should in the same way increase it if the next inspection is successfull?
A rule must be set on how many failures in a row must trigger a calibration frecuency change. A similar rule should also exist for the case of instrument allways meeting the calibration inspection, shall we risk to increase the calibration interval if the instrument has been within the calibration tolerance during the past 5, 10 inspections?

I would say that once a year the calibration history be analyzed and frequencies adjusted if applicable.

Josh, Darth,
The sample we use for analysis must contain items that are
- identical (make, model, size, rating, same operating context
- independent, i.e any item in the sample does not influence the performance of other items
If these conditions are met, we can pool failure data and compute MTBF or failure rates(lambda).
In practice we make approximations because it is very hard to find, in field data, these conditions to be fully met. So we accept largely similar items in similar duty, e.g., a set of 100 pumps are unlikely to be identical in make, model, size, rating or operating context. If we insisted on these conditions strictly, we will end up with a sample size of 1, which gives us pooor confidence in the results.
Once we have MTBF data and the required availability, as in Josh's case, yes we can determine the test intervals.
Josh I have written to you separately; there is an error in the data used in your MTBF calculation but the rest is OK. Just correct the MTBF value and you should be OK. Your sample size looks a bit small to give great confidence, but that will improve as time goes by and you collect more data.At tjhis stage, I would take a conservative view and adjust the test intervals caklculated upwards, at least slightly
Age exploration methods are useful when sample sizes are very small. .
I would take manufacture's recommendations with care; they will be very conservative because they dont know your operating context and have to provide for worst case scenarios.

Thanks Vee.

I can gather more data from other fields to increase the confidence level.

What is the confidence level we should have? Is 90% ok? Is there a way to calculate this?

How much data should be analyzed to have an appropriate confidence level?

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

Josh,
90% confidence limits are OK.
See page 216-218 of Robert Hansen's excellent book on Overall Equipment Effectiveness, Ind. Press, NY, ISBN 0-8311-3138-1, or books on Statistics, see Chi-Square test.

I have never seen using a confidence level below 90%. Mostly 90% and 95%.

Darth,
One of the problems gathering field data is volume. The higher the confidence you seek, the bigger the data set you need. If you are a scientist in a laboratory this may be relatively easy, as you have control on the 'independant' and 'identical' criteria.
In the field however, this is difficult, If for example, you want to maesure Fail-to-danger rates of Pressure Relief Valves, does your sample contain
- All PRVs in the Plant?
- Only steam PRVs?
- Only steam PRVs working at 20 barg?
- Only steam PRVs working at 20 barg, balanced bellows type?
- Only steam PRVs working at 20 barg, balanced bellows type, 4" 300# x 6" 150# flanges?
- Only steam PRVs working at 20 barg, balanced bellows type, 4" 300# x 6" 150# flanges, make Crosby?

We can go on disecting the set till we have only one valve left. What do we do with our confidennce band then?
There is thus always a compromise between the desire to have high confidence (i.e., large samples) and meeting the two statistically rigorous conditions I mentioned earlier. By the way, larger samples cost more to test as well.

The lack of data or lack of a sample size large enough to validate data is a common concern I hear. While it's possible to use a Weibull analysis in many situations that would reduce the need for larger sample sizes, it makes me wonder if people (you people in plants) would be willing to contribute to a database of failures (not sure if failures is the right word) of various types of equipment that would allow everyone to accumulate this type of data?


Vee, would such a database be possible collected from many different sources?

Everyone else, what would be your concerns about contributing to such a database?

Joe,
Of course it is possible to collect failure data from a number of facilities for analysis. The offshore O&G industry do so with OREDA, and I believe the Chemical Manufactures Association do something similar. The National Power 10-yearly surveys in the power industry, Edison's surveys etc are all pooled sources. IEEE produces data bases using a similar approach, also RAC in Rome(Non-electrical Parts Reliability Data).

As you know, confidence is not not merely based on the statistical sample size. We must know that the items are substantially identical and operating in the same or very similar operating context. For this, we need a taxonomy, for which we can use ISO 14224, based on OREDA.

Industry groups can pool data, but it needs a lot of organization and coordination. Lots of problems crop up, e.g., CMMS systems which are configured differently, different defintions used etc.

Even within Companies, many definitions and protocols exist which make it difficult to collect and pool data, e.g., operating contexts may differ from site to site etc.

In sum, some people have got over the hurdle, others are still struggling. The top performers somehow manage to collect data and USE them

Joe, it would be good if you can get us the links to those sources mentioned by Vee. Fast quick results to get and maybe we can built on existing efforts rather than starting from scratch.