Capability Study on a New Machining Process - Issues

Miner

Forum Moderator
Leader
Admin
It is too soon to jump to that conclusion brahmaiah. As Bev D correctly observed, the within sheet variation is the largest source of variation and should be minimized first. If you subgroup as you recommend before reducing this variation, the subgroup scheme will mask this variation.

mayor should first focus on reducing the within sheet variation. Once all locations are statistically equivalent, mayor should then decide on a rational sub-grouping scheme. After all points are statistically equivalent, there should be no need to average these points. A representative point should be selected. mayor can then decide whether an Xbar/R, Xbar/s, IMR, or other chart will work best.
 
B

brahmaiah

It is too soon to jump to that conclusion brahmaiah. As Bev D correctly observed, the within sheet variation is the largest source of variation and should be minimized first. If you subgroup as you recommend before reducing this variation, the subgroup scheme will mask this variation.

mayor should first focus on reducing the within sheet variation. Once all locations are statistically equivalent, mayor should then decide on a rational sub-grouping scheme. After all points are statistically equivalent, there should be no need to average these points. A representative point should be selected. mayor can then decide whether an Xbar/R, Xbar/s, IMR, or other chart will work best.
Well you have concurred with my view that assignable causes of vaiation should be eliminated first.
I donot agree with feeding 8 readings for one charactristic of a part.Thickness of a sheet in one characteristic.If you measure one characteristic of a part in eight spots you will be measuring the taper.We are interested to know wether the part is conforming or non-conforming.when asked what is the thickness of this sheet can we give 8 answers. All 8 measurements represent one characteristic of the product. A characteristic cannot be partly conforming & partly
non-conforming.
In a control control chart a data should represent one characteristic fully and not 8 of them together represent one characteristic.At the end of measurement we have to answer how many parts are accepted and not how many spots are accepted.
V.J.Brahmaiah
 
Last edited by a moderator:
M

mayor

I agree, too much variation within. I have ID the cause, vacuum table not flat, and I'm working on a solution now. In regards to subgrouping, why would I not want to subgroup equal to one sheet's data. I understand the 8 measurements are measuring the same thing, thickness, but that is what help ID the problem within the sheet.
 

Bev D

Heretical Statistician
Leader
Super Moderator
So is it correct to say that the process is currently not capable? Hopefully after correcting the table flatness the process will be improved and be capable.

Thanks for the feed back.

Yes - you are not capable.
even if you correct the within sheet variation - your latter sheeets were all centered in the lower half of your spec range. Of course, once you have the within sheet varaition reduced it may be a simple matter for you to center your process :)

one aside: if you are really looking for the "capability" of your process in terms of it's current ability to meet spec, you should be using Ppk or "long term Cpk. (the term Cpk is often loosely used. most Customers mean Ppk when they say Cpk. In fact JMP - minitab's primary competitor - used the terms long term Cpk and short term Cpk :confused:. If you are looking for potential capability calculating Cp or Cpk is dangerous unless you actually understand the subgrouping scheme as in your example)
 

Bev D

Heretical Statistician
Leader
Super Moderator
Well you have concurred with my view that assignable causes of vaiation should be eliminated first.
I donot agree with feeding 8 readings for one charactristic of a part.Thickness of a sheet in one characteristic.If you measure one characteristic of a part in eight spots you will be measuring the taper.We are interested to know wether the part is conforming or non-conforming.when asked what is the thickness of this sheet can we give 8 answers. All 8 measurements represent one characteristic of the product. A characteristic cannot be partly conforming & partly
non-conforming.
In a control control chart a data should represent one characteristic fully and not 8 of them together represent one characteristic.At the end of measurement we have to answer how many parts are accepted and not how many spots are accepted.
V.J.Brahmaiah


well it depends on the physics...first it's not clear to me that his within sheet variation is measurement error induced by 'the table' or if 'the table' is inducing actual thickness variation. Then we must be concerned with the use of the sheet. It could be a non conforming sheet if any one location is out of tolerance. It will depend on the use. simply averaging the within piece variation to get a single answer can be very dangerous: measurement must meet intended use. I had a problem many years ago where the CMM programer averaged 3 within piece readings to "average out what he thought was emasurement error". OOPS. the readings were located in psoitions wher other parts of an assembly were connected. the thicker positions caused stress on the otehr parts eadign to catastrophic failure of the entire assembly. Lives were actuall at risk although the failures were controlled and we only ended up paying millions of dollars in warranty and consessions.
 

bobdoering

Stop X-bar/R Madness!!
Trusted Information Resource
I agree, too much variation within. I have ID the cause, vacuum table not flat, and I'm working on a solution now. In regards to subgrouping, why would I not want to subgroup equal to one sheet's data. I understand the 8 measurements are measuring the same thing, thickness, but that is what help ID the problem within the sheet.

The problem with subgrouping data from one sheet is you are running into a quandary found also in web processes. In a web process you can sample across the web or down the web. Down the web represents the variation in terms of time, and therefore the range represents variation over time - which is what you are really trying to determine to make the data predictable.

Sampling across the web gives you an understanding of the variation of side- to-side issues, but the range is not a time function. Using it to develop the control limits - especially - will create problems in interpretation. For discrete part data (one data point per part) which is the typical and most correct use of the data, the range is a function of time to one point on the part. I doesn't hurt to know within part variation, or track it something similar to a X hi/lo-R chart where the range will be parallelism variation in your example. More variation in parallelism can be evidence of a an issue (such as vacuum leak.)

So, you run into a similar problem if you try to calculate control limits based on within part rather than between part. Multi spindle machines have the same problem (screw machines for example). Better to track the worst spot across time - or all spots separately across time, if need be.
 

Bev D

Heretical Statistician
Leader
Super Moderator
An alternative approach which I have found to very useful is to continue to measure your 8 locations - or maybe fewer if you can capture your within sheet variation with fewer ocations - and have two separate charts:
  1. A chart on the sheet average with the control limits calculated using the Moving Range of the averages
  2. A Range or Standard Deviation chart for the within sheet variation
This way you see both time variation in the sheet itself AND any chges to the within sheet variation. Since they have seperate causal systems it makes sense to chart them seperately...

The only way to know what will work for you is to try a few different approaches...
 

bobdoering

Stop X-bar/R Madness!!
Trusted Information Resource
I see the same thing and have indicated our vacuum table with results that match the data. Looks like I will be flattening the table. I use Minitab, so the Cpk formula used is min(CPU,CPL), which uses the within sigma. Spec limit is +/-0.0005".

I also looked at the control chart (I-MR) for each location and they are out of control as well, so there is variation sheet to sheet, but I agree that within is larger.

Just a curiosity - how consistent is the raw material thickness and parallelism?
 

bobdoering

Stop X-bar/R Madness!!
Trusted Information Resource
Not very - thickness is 0.046+/-0.004. We don't check parallelism.

That can be very important when trying to control to a spec limit of +/-0.0005". You may have to do roughing/finishing to blank the material closer to allow the finishing to keep such a tight tolerance. That is a very common scheme in metals, as variation in blank size will change tool pressure, etc. in the cutting process - causing (or amplifying) variation, too.

The other option is buying tighter tolerance materials (in metal terms, ground stock).
 
Top Bottom