MSA (measurement systems analysis) on several CMM's - Correlation issue

[Scary Bird]

Good day to you all

Here’s my problem: We currently use 2 CMM’s on site to measure aerospace fan blades. We then ship the blade to our customer at another location; they are also part of our company. As the foot of the blade is a critical area to them they re-measure the blade on their CMM. Our results are no were near theirs, so we have a correlation issue. None of the CMM’s in question have had an MSA conducted. What’s the best way to proceed? Also this is a new product with its own unique challenges.

Conducting an MSA on 3 CMM’s, taking 10 blades measured twice by two operators for each machine would be a huge resource issue, and for two of the machines we would be able to have the machine for that length of time.

Another factor to throw into the pot is the way the blade is located on the CMM, our site and our sister site (customer) are different.

One person has suggested that we take an artefact and each site measures it a couple of times per operator and then analyse the data using nested Anova. Would this show repeatability and reproducibility to a statistically significant level to allow us to have confidence in the results?

Really, what would be the best way to proceed without doing a full-blown MSA on every CMM, but giving us the confidence in the results of the machine?

Thanks for your assistance.

 

[Claes Gefvenberg]

As the foot of the blade is a critical area to them they re-measure the blade on their CMM. Our results are no were near theirs, so we have a correlation issue.

.....

Really, what would be the best way to proceed without doing a full-blown MSA on every CMM, but giving us the confidence in the results of the machine?

I see your problem. Let's start with some questions:

Could you providee us with some data on what is being measured and how?

Are the CMM's in question identical? Model, manufacturer, software, and so on... They all have their quirks, particularly when it comes to software. Machines from different manufacturers may very well evaluate identical data differently, and in some cases even incorrectly. Manufacturers do cut corners when they develop software, and they may be using different formulas. This is particularly evident in the case of form and position tolerances.

The operators present another obvious potential for differing results. They may measure the same thing in two different (but correct) ways and come up with different results. The way the blade is located on the CMM is one example of such differences. Have the operators been able to meet and discuss this? There is a fair chance that that would sole your problem.

Could the transport affect the product? It's been known to happen...

/Claes

 

[David Hartman]

Do your drawings use GD&T dimensional standards (it seems to me that this would limit some of the variances in measurement methods)?

 

[bmccabe - 2006]

We have the same issues here. My solution is that we make a surveillance master and perform an anova on it, from each machine. I set the tolerance for the measurements at 1.3X (allowing for a 6sigma curve to fit therein) the OEM’s published performance specs.

[Here we measure such things from a program, and have concluded the operator has no influence on the performance of the gage by pressing the run button. I mention this because the anova will attribute some error to operator – And this variation is actually a “form error”, on the part.]

If the anova fails, the field tech is called to calibrate the machine. If it passes – We establish processes capability by re-measuring the artifact (surveillance master) at regular time intervals, and plotting the results on a pre-control chart. The spec limits for the pre-control chart were set at 1.7X OEM specs (allowing for process drift, and reducing the tendency to over-react to normal variation.).

When the control chart violates one of the “rules”, we re-calibrate. In the meanwhile, we document the systems ability to perform within specifications, and archive the data in the SPC database.

I know this seems like alot of work, but so far, it's proven robust. Hope it helps.

r/b