Hi there,

I have been cracking my head on this and could not figure a proper way to conduct a GR&R on the dual axis laser gages that is use to track the diameter of extruded plastic profiles. According to MSA Manual on page 143, I'd consider this a non-replicable measurement system. Problem is, when I use minitab to conduct the GR&R, i got a rather unacceptable %GR&R of 100. Another factor to consider is that the extruded profile has a rough surface which is absolutely normal. Please refer to GR&R result below:

Gage R&R

%Contribution
Source VarComp (of VarComp)
Total Gage R&R 0.0000000 100.00
Repeatability 0.0000000 97.42
Reproducibility 0.0000000 2.58
Part-To-Part 0.0000000 0.00
Total Variation 0.0000000 100.00


Study Var %Study Var
Source StdDev (SD) (6 * SD) (%SV)
Total Gage R&R 0.0000905 0.0005429 100.00
Repeatability 0.0000893 0.0005358 98.70
Reproducibility 0.0000145 0.0000872 16.06
Part-To-Part 0.0000000 0.0000000 0.00
Total Variation 0.0000905 0.0005429 100.00


Number of Distinct Categories = 1 :nope:

If somebody look at the result, they might say our measurement system needs serious attention but in fact, we are making good parts! 98.7% of variation contributed by repeatability because the laser gages is not measuring the same parts during the study! CPK has been always above 10. I'd really appreciate it if someone could shed some light on this. Thank you in advance.

Please post your data, and describe how you collected the samples that represent the parts.

see attached.

sample intervals: 10 seconds.

The raw data attached is only a 1 hour data with cp, cpk and sigma.

Edit: LSL and USL are 0.100 - 0.110"
Target: 0.105"

It looks like you are force fitting capability study data into an MSA. The samples collected for an MSA must be planned and sampled specifically for an MSA.

In an extrusion process, you will typically see some degree of autocorrelation. You should analyze this and determine the period for the autocorrelation. The parts selected to represent each part must come from different time periods such that there is no autocorrelation between "parts". This is probably why your part variation is zero.

Given your comment about the roughness or ripple in the extrusion, this is variation in form similar to shaft roundness. You need to develop a consistent method for dealing with it, such as averaging, or max readings.

Wow, thanks for the tips on sampling. Problem solved. Thanks!

I ran an autocorrelation analysis on your data. It is definitely autocorrelated at the 10 second time intervals. The analysis shows that the autocorrelation is significant out to ten lags. If you increase the time period between measurements to 1 minute 30 seconds you can eliminate the autocorrelation.

The variation within the 1 min, 30 sec window is a combination of the gage repeatability and the variation in the rough surface.

In a previous life, I worked with laser measurement of rubber extrusions. We could measure the roughness of the surface and detect a sine wave pattern caused by the flights of the extruder screw as it rotated. We ended up scanning a 12-inch section of extrusion at a fast sampling rate and averaging the data to obtain a single measurement. We then took the next sample at a time period where there was no longer autocorrelation.

I ran an autocorrelation analysis on your data. It is definitely autocorrelated at the 10 second time intervals. The analysis shows that the autocorrelation is significant out to ten lags. If you increase the time period between measurements to 1 minute 30 seconds you can eliminate the autocorrelation.

The variation within the 1 min, 30 sec window is a combination of the gage repeatability and the variation in the rough surface.

In a previous life, I worked with laser measurement of rubber extrusions. We could measure the roughness of the surface and detect a sine wave pattern caused by the flights of the extruder screw as it rotated. We ended up scanning a 12-inch section of extrusion at a fast sampling rate and averaging the data to obtain a single measurement. We then took the next sample at a time period where there was no longer autocorrelation.

Autocorrelation is something new to me. Could you please tell me how 1 min 30 sec is obtained from your analysis??? I might give it a shot. Thanks.

Refer to the chart attached to my last post. When the blue bars are outside the red limits, a significant autocorrelation exists.

In your data, this autocorrelation continued until the tenth lag. A lag is the time period between successive measurement. Ten lags corresponded to 1 min, 30 sec in your data. Note: You can extend this time, but don't sample at a less frequent time period.

Miner, I did an autocorrelation to a different set of data. Could you please help interpret the result? Thanks in advance.

In this study, the process is showing significant autocorrelation out to and including 11 lags. The 12th lag is no longer autocorrelated.

This means that if you take your time between samples and multiply it by 12, you need to wait this long between measurements to see actual process variation.

Miner, I've researched further into sampling data and I'm actually seeing what you've mentioned earlier. A sine pattern! Correct me if I'm wrong, in order to get rid of the autocorrelation, sample interval should be increased 18 times. Current settings are

sample between interval = 12 seconds
samples per subgroup = 5

Due to the nature of our process, we are collecting 1 subgroup per minute for 60 minutes until SPC chart is reset. As mentioned earlier, I/MR charting is not available. I've tried 2 samples per subgroup with 30 seconds interval but autocorrelation still exsit. What changes would you recommend based on my current situation to get rid of autocorrelation?

Attached is the autocorrelation function. Your input is greatly appreciated. Thank you.

Good to hear back from you.

You are correct. Based on this, 18 lags x 12 s/lag = 216 s = 3.6 minutes between samples should eliminate the autocorrelation. Remember that you can always go beyond this to 5 or 10 minutes.