Hi,
in a recent discussion with a customer we have recognized some things in the Reliability Forecast and Life Distribution Platform I could not answer right away and which where somehow surprising or an enhancement.
1. Reliability forecast:
- Data is in nevada Format and organized in a way that several identifiers (groups) are listed, so V1_0 with data in Nevada Format, V1_1 and V1_2 beyond also in Nevada Format. (see attached dataset Beispiel1_Anonymized.jmp)
- If I do a reliability forecast with data filter set to V1_0 or V1_2 (V1_1 we can ignore for this example) and use Quantity as Prod Count, P-month as Timestamp, and the date columns as Failure Count, it works nicely for V1_0 alone, V1_2 only and even for both V1_0 and V1_2 both selected. But as soon I add X_1 column (which contains the V1_0 and V1_2 nominal values) as a group ID (picture below) it fails to work and brings an error message data format is invalid.
- I thought the Goup ID will allow for having in one window the same analysis once for V1_0 and once for V1_2 (assuming data fiter is set correctly ) - where does the error come from?
- Reliability Forecast:
- Sometimes there is a repair right in the first month of service or even before in service there is already a repair. So a 1 or other positive number on the diagonal or in th elower triangle. This results in a reduced set of possible distributiions (e.g. no weibull). How to deal with this? How can one workarund to still be able to use weibull or other better fitting distributions? (Just use small production and change the 4th row of column 09/2009 or 10/2009 to 1 and run the script to see the difference)
- Life Distribution: (see Beispiel2 Anomized.jmp)
- if you have repairs before install date this will result in negative values. How to deal with that? Also when you have competing causes? (is adding an offset a realistic solution? or does this have bad impact as well? Another option could be to use the production date, but ths would not reflect the correct forcast for the install date, I'd assume)
- if you have repairs e.g. few days after install date but want to look on a monthly base this would result to zero time. How to deal with that?
- Sometimes a Fit Mixture is more accurate than the Weibull or other distribution itself. How to get the mixture fit in comparison to the original plot of the distributions in one window? I know one could copy the graph content, but it would be much better if there would be an option to include the mixture by option to the list of other distributions to show in the original plot. Also when you have competing causes there is o way to have the mixture fit to be selected. This would be a great advantage said my customer. Are there any plans or is this hard to achieve by a workaround? (I know this is more a question for the developers)
- For all distributions you can use the option Fit survival, however, this option is not available for the mixture fit. Can this be included to the platform? Is there a simple workaround?
Thanks for any feedback on those questions.
Martin
Hi Peng, thanks so far. I'll investigate further to see if the workarounds will work for the customer so far and to understand what the business impact would be for the zero inflated case. Will add some notes here soon.
Martin
Hello Martin,
The reported problem is a bug. I have fixed it in JMP13. To work around in the current releases, deleted the group, if you exclude all rows in a group.
For Reliability Forecast question, the time points are all singleton, e.g. 09/2009 means the first tick in the first day in Sept in 2009. Therefore, you need to differentiate the row timestamps and column timestamps to let the software know there is a beginning and there is an end. Failures before the beginning is an error.
You probably don’t want to play the odds to put failures right on the beginning. You will not get desired results. Because failures at the beginning means failure at time zero. A ZI distribution is required to analyze such data, and ZI distributions are not supported in the platform. Zero-inflation failures are rare. If you do need it in Reliability Forecast, please file a request.
For the first two questions about Beispiel2 Anomized.jmp, I guess you are still talking about Forecast platform. I think the first one is an input error, no doubt. The second one, you have play with the row and column time stamps and aggregate new counts to get desired time intervals. The Nevada format is popular, because it is already in the desired interval for forecasting, produced by IT.
For overlaying mixture with individual, an alternative, is to fit single-component mixture. E.g. fit a 2-Weibull mixture, and fit a 1-Weibull mixture.
For survival curves, try to revert Axis min-max, a check box in the Axis dialog. Somehow, the CI shade is lost, I need to investigate. I guess not trivial.
Thanks,
Peng
