Asking advice - lowering the frequency of tests for our products

[#1]

My company is new with JMP, I'm now working with JMP for a year.

With selfstudie I try to get better with it.

At this moment I'm trying to map out the correlations and influences of our current data and predict what the outcome can be if we let out/exclude some of the current experiments. 

We lose a lot of time by testing. We have around 12-18 test per batch. (not including the typical test/extern test we are unable to do, that are done once in a couple years per product) 

And the main goal is to lower the frequency of the test rate. 

At the moment, we are testing our products on three levels. 

Level 1 - First check up (done with all the batches, things can be added/changed by the opperator, before the end product is made)

Level 2 - Standard test (done with all the batches, test done with the end product)

Level 3 - Extended test (doing on frequency, mostley done once every 5th batches/charges.)

If something goes out of spec/control limits at level 1 or level 2, level 3 will be done no matter if it was the 3rd or 4th batch instead of the 5th. 

Level 3 is kind of leading, as long as those are in the spec limits, then level 1 and level 2 are ok. 

So you can see level 1 and level 2 more as control limits, do we have control of our product? 

Is it possible to say with this number(statistic), those test are stable, those test are less stable.(important/less important)

The frequence of every 5th batch can go to a higher number depending on the stability of the product.(low variety)

Have anyone advice which function to use in JMP? 

I'm already thinking of:

- Multivariate

- Variability / attribute Gauge

[#2]

Here are my thoughts.  It seems it would be best if you could understand what are the factors (X's) that affect the performance measures of your batches (Y's).  You could design sampling plans to discover where the most variation is (e.g., measurement, within batch, batch-to-batch, changing lots of raw ingredients, etc.) or you could design experiments to manipulate the variables that could impact the "quality" of the batches.  JMP can help do both of these.  Once you understand the relationships between the X's (input variables) and the Y's (output measures), you can predict the batch performance based on your knowledge of those relationships (and develop prediction models in JMP).

You are relying on chance (probability) if you just look at the outputs and compare them to spec. and react to those comparisons.

[#3]

Thanks for your reply!

I understand were you want to go.

I will try to make a plan. If I can't figure it out, you will see me here again

[#4]

Establishing true cause and effect is the holy grail, and it can be a long (sometimes difficult!) road. But, from a pragmatic point of view, if you have lots of data you may be able to exploit what you already have, at least for the purposes of establishing any additional risks associated with taking fewer measurements (which seems to be at least part of your intention). To do this, you would need to study the patterns of variation in your existing data and then, through simulation, look at what operational decisions would have made differently if you had used a more relaxed sampling scheme.

The general idea is that, early on (when you don't know much) you 'oversample' to understand the empirical patterns of variation, then as you accumulate more evidence over time, you can move to a more appropriate sampling plan. But note that this approach would never alert you to the fact that you actually need to take more measurements. Stating the obvious, but it all comes down to how you acquire and exploit process and product knowledge.