GRR (Gage R&R) Method for Camera Vision Inspection System Hardware and Software

[orion - 2004]

I am developing a program to simulate Gage R&R calculations to be used in the automation field. The problem is, my measurement system consist of a camera and a vision software (camera grabs images, vision software take measurements), so actually there is only one VIRTUAL APPRAISER.
Currently, my program runs a 10 Parts - 3 Trials - 3 Appraisers pattern, but I think it is not appropriate and illogical.
Is there any waysI can modify the GRR to suit my situation, and is there other equivalent method that I can use, and is comparative to the present GRR method used in the industry?
Thanks a lot.

[bpritts]

You are right that the standard GRR tests do not always make sense in more specialized, e.g. automated testing, environments. First, though, I would recommend the recent automotive GRR manual from the North American industry if it's available to you -- Measurement System Analysis, Third Edition. Check aiag.com for availability locally. The third edition has a lot of good thoughts on automated testing (big improvement over the 2nd edition.)

Maybe I am just stating the obvious but a few quick thoughts.

1. The instrument repeatability obviously needs to be retained; 3 trials per part would do this.

2. The "appraiser" side - this depends on issues like part presentation. If there is any possibility of operator sensitivity then you should still do multiple operators. How is the part positioned for the vision system? Is this operator sensitive?

3. More important than the standard GRR is really the correlation between your instrument's calls and the "truth" to the extent you can measure it independently. For example, if you are taking dimensions you can measure the parts independently using traditional means, then verify the correlation between the independent (assumed "true" measurements) and your device. Look for type I and type II errors where parts pre-determined to be good are rejected by the device and vice versa.

Depending on the nature of the inspection, you may want a much larger sample size than 10 parts to do this.

Anyway, there are a couple of thoughts.

Best regards, Brad

[orion - 2004]

Firstly, thanks for your reply.
Actually, I have purchased the MSA 3rd Edition. I read through it and found nothing to be suitable for me (or maybe I missed something...?) except for Appendix D, MSA 3rd Ed. In it, it did explain briefly on using one Appraiser as a quick GRR test, but it is not recommended even by the book. Plus, it is a summary of method, so I do not really know how to implement them. Perhaps someone could explain to me?

On your thought about the Appraiser part, the one that is doind the measurements is the Vision Software, and in my opinion, there should not be any operator sensitivity that involved. Of course, I may be wrong, but let's just assume that there is only one Appraiser.

I would appreciate more feedbacks.
Thanks

[bpritts]

To guide the discussion a bit, could you add some more detail about the vision system usage. Is it doing attribute measurement (e.g. looking for component presence/position on an assembly); dimensional/variable measurement; both?

I will go along with the assumption on no operator variability.

Then look at:

"Accuracy" -- feed in a number of samples of known quality (good or bad); does the device correlate with the known quality?

"Repeatability" Try repeating the measurements (e.g. 10 parts, 5 trials). Still get the same results?

"Stability" = do the same, or abbreviated, test every day to see if the results stay the same.

If you tell a bit more about the process we can discuss more..

Brad

[orion - 2004]

Good to hear from u again,Brad.
I will answer your questions in the order as they appear.

The Vision System can do both attribute (checking for presence..? ) and variable (checking for dimensions..?) measurements. Actually, I do not understand what is attribute abd variable measurements, care to explain ?

"Accuracy" - yes, it does correlate with the known quality, limited to the camera resolution.

"Repeatability" - If the sample is stationary at its position (no movement whatsoever) during each measurements (grabbing of image, and processing) then the repeatability will be good. If the sample is took out and put back in, the results will vary a little.

"Stability" - The vision system is quite stable, and will produce almost the same result in a range of time, if the sample is non-destructive over time.

A brief explanation on the Vision System. Basically, a manufacturing unit will be moved (using conveyor belt for example) to a position where the camera can grab a good image. Then the unit stops at the position, and the camera will grab the image, and let the software "measure" the units (length,width, area etc). Then, the unit will be moved out of position, and the next unit will come in, and the cycle just continue.

Currently, to simulate a 10 Parts-3 Operators-3 Trials GRR, I use the following method.
Part 1 will go into position, measurements taken, then Part 2, then part 3 till part 10. Then a new cycle begins, with Part 1 at the helm, then part 2 and continue to Part 10. I will do a total of 9 cycles, and thus each part will have 9 measurements (a virtual 3 Oper * 3 Trials).
There it is, a 10-3-3 pattern. But if we look carefully, there is only one Operator, with 9 trials and 10 parts.
I hope that my long and winding text would not bore you away.
I would really appreciate feedbacks.

Regards

[joecebo]

I too am working on something I believe to be similar and need assistance, perhaps you can help? My robotic system applies an adhesive inside a groove of a rubber seal. The adhesive contains a chemical which enables the robot's photo system to determine if there is sufficient adhesive applied to the seal. If, after the machine reviews the photo and determines there is sufficient adhesive, it attaches it to a mating part; otherwise, it rejects the seal and continues by picking up the next seal and proceeding accordingly. This is all done very quickly, within seconds.
I am attempting to do an attribute Gage R&R and do not know where to begin. The camera makes the decision to accept/reject the seal and I have no idea how "good" our measuring system is. Can/would you help me?
Thanks,
Joe

[statdoug]

You mention that there is variation if the part is removved and replaced. Is this caused by variation in the presentation of the part? If so, it is really a part of the measurement system and needs to be tested. The variation seen in the measurements of the part when it is re-imaged and the variation seen in the part when it is measured from 1 image is the first aspect that I would want to look at. The second would be how the variation when the part is re-presented by the conveyor compares to the variation from the single image and from the re-imaged part.
Probably for the truist representation of gage error in this application, the re-presented variation should be viewed as the true Equipment variation, and Appraiser variation should be re-runs of the parts re-presented multiple times, run on different days.

[statdoug]

Is it possible to present the same part to the camera multiple times, or does the part automatically get assembled if it is good?