MSA when neither Nested nor Crossed Gage R&R is Applicable

MSA in destructive mechanical testing (tensile). We have to qualify and validate all of our tests. We went for Nested GRR but we failed in creating the (we went for 6) series of homogenous parts. GRR nested (minitab) returned a NDC of only 2, Total GRR 18,5%. I guess that part-part variation was too large compared to the averages between series. Oh well what do?.
How about this for a MSA when high Cp:
Calibrate load cell and tensile tester.
Get 3 Trained operators. Name the trainer of other operators the master.
Run 3 sets of 50 parts all at one time. From single cavity or machine as good as you can.
Assume parts to be identical across the 3 sets and little to no part-part variation within set.
Test the 3 series using the 3 operators.
Pass normality test for each set
Run ANOVA. If good then there is no significant operator contribution.
The master operator should have the "perfect" average so add the distance of means from your poorest operator to your acceptance criteria. Get the pooled standard deviation and multiply it with 6. Add this to your acceptance criteria.
Pass qualification test with Cpk of 1,67.
This can from time to time triple your original acceptance criteria so this is best for very high Cp product, so is there a better more realistic way? All your part-part variation is converted into measurement uncertainty

Thanks

Re: MSA when not Nested or Crossed GRR is Applicable

Read Understanding Sources of Variation in Mechanical Testing Results from Instron.

This article describes the use of springs in lieu of actual parts to provide a replicable R&R study. In our situation, we found springs with rates that would provide results similar to the destructive tests that we performed. We performed a crossed study on the springs and compared the measurement variation to historical part variation.

Re: MSA when not Nested or Crossed GRR is Applicable

Thanks for your reply - it is mostly appreciated. We did something similar to your spring setup. We need the detection of break as a part of the qualification of the tensile test machine since it is not a part of the calibration of the load cell and tensile testing machine. Therefore we introduced a second load cell with higher sampling rate and resolution - also calibrated - and it has smaller uncertainty per certificate than the load cell for the tensile testing machine. This is mounted under the bottom fixture. Next we have an array of magnets of different strengths which we pull from each other and map the difference between the two load cells within the range of the load cell for the tensile testing machine for bias and linearity and %R&R. For the specific test we will perform we add this difference to the acceptance criteria. The machine is then qualified. We now take 3 operators who test 50 parts each. We strive to have the full process variation represented in each operators batch. We run anova to test for operator dependability. If none then good we validate the test by passing the elevated accept criteria by cpk of 1,67. However in some of our products we have so little variation that we cannot pass the anova eventough averages are witin 0,05N from eachother. Would it be in your opinion sufficient to loose the anova test and accept some difference between operators and pass the test with cpk 1,67.

Or are there other options like P/T ratio with global mean and standard deviation to qualify the test with? Are there ways to compare the standard deviations for each operator not focusing on the means like anova. I am not fond of XBar and R but what could the the acceptance criteria be if we went for that?

Depending on the resolution required, I used fishing line to do gage R&R for lighter pull strengths (approx 6 to 12 lbs). It is fairly repeatable, is available in several pound test strengths. You may have to tie it to something that fits in the grips, or it will slip like crazy. Check here to get the correct knot to keep the line from slipping. I have tied it to eye-hooks.

Another point is are you testing to ultimate strength (failure)? That is usually the least repeatable for any material. Testing to yield strength is more consistent, if you can chart the curve.

We test to failure. There is no good point of yield to detect on the parts we test. The standard we test to specifies a force and direction without failure for 15 seconds. This is performed in the production with x parts each hour but as for pass-fail tests goes there is not a lot of production variation knowledge to gain. We use the force as acceptance criteria but pull until failure. Pass the acceptance criteria by 1,67 for qual testing.
MSA is a four step process at my workplace
1. calibration
2. qualify test machinery
3. Use DOE to eliminate variations in specific test
4. Qualify test by GRR or other method currently unknown.
Regarding our magnet-dual loadcell setup. we get really good repeat and reprod for failure detection and proof of accuracy so we can qual the test machinery just fine adding the found uncertainty to the acceptance criteria. However this is only half of it. We need to do it with actual parts and 3 operators aswell using the actual setup to prove operator indenpendency and precision. Here it gets tougher since we either failed to do a nested GR&R or have too large variation do a P/T ratio or have too small variation to pass an anova test. I really need one method for all of our test method qualifications.

As I read your reply again and miners aswell it seems as if we take it a step too far by mixing the actual test into the MSA somehow.

Again thanks for your reply

2wrongsmake1right said:

As I read your reply again and miners as well it seems as if we take it a step too far by mixing the actual test into the MSA somehow.

Click to expand...

The point of Gage R&R is to determine if the gage is "the right gage for the job" and provides adequate statistically significant resolution. If you can show that, you are doing good.

Tensile testing is not a "gage", and tensile strength is a terrible thing to try to test to begin with, because there is so much natural variation in the outcome. So, that is why trying to determine if the gage is good when the output is so variable because of the sample and its physics is such a difficult issue. Without consistent samples, the gage is blamed for variation that it does not contribute.

I couldn't agree more, however I am still left with the requirement from RA/QA to verify all contributing variations to any test result whatever the test, even the really old products we still produce. Therefore I would much rather perform the GRR with our simulation rig which I know we can pass. Qualify the tensile tester as an adequate "gage" and let 3 operators test 30-50 parts each and if the entire population is normally distributed then pass the acceptance criteria by cpk 1,67. I can to some degree decide myself how we as a company are going to do it as long as I can statistically prove all contributions to variation all options are open as there is no real GMP for this.