Design of Experiments to Reduce Variation in a Lathe Operation

[Coleman Donnelly]

I have a project that I am working on to reduce variation in a Lathe operation.

I have about a dozen or so features that are being inspected in process and I have boiled down my control variables to Spindle Speed, feed rate and fixturing pressure (vacuum pressure).

So I am trying to set-up my study so I can establish the number of samples that will be needed.

The effect that I am looking for is to reduce common cause variation from the process. By reducing the overall variation of the process (or sigma value) I can improve the CPK, reduce scrap rates and potentially reduce my inspection frequency accordingly since the process inspection results will be using control charts to control the process.

Back to my problem:

I have 3 variables that I can play with which gives me 8 corner points. I believe that I want to use center points as well in order to improve the vissibility of curvature of the settings.

I also heard something about star points, I don't know what they are or when they are required/applicable...

So this means that one run of my DOE=9 samples

Now I need to figure out how many replicates and repeats I am going to want to do.

If I do a power and sample size study I need to plug in 3 factors and it will give me the 4th...

Number of replicates
Effect
Power
Center points

So -

Center Points -1
Power - ?
Effect - ?
Number of replicates - ?

I am assuming that Power should be between 80 and 90%, but I am unsure of this. My perfectionist brain is telling me that I should try to get close to 100%, but I have been cautioned that if Power is to high than I will get a false signal in my design.

so I run power at both 80 and 90

Effect. I am trying to improve my process Cp/Pp. If my current process state yields a Cp of 1.15 and my target is 1.66 does this mean that my effect response I am looking for should be set to 0.51? It seems that any effect in the right direction is a good thing...

So I am not sure what to put here.

I have:

center points 1
Power 80, 90%
Effect 0.10
Replicates?

So here is the problem... Typically I would run a 30pc short run Cpk study to evaluate the Cp of a given characteristic of my part... This takes my original DOE sample size from 9 to 270 with 0 repeats and 0 replicates... this seems to me like a bit much...

So now I back up a step:

If I can calculate my sigma from an initial capability study for each of my characteristics than cant I determine a size smaller than 30 to establish my Cp/Cpk within a reasonable 95% confidence interval. Can I than reduce my 30 pc requirement to the largest requirement to have 95% confidence that my returned sigma value from the DOE is stable?

This is my idea, but I don't exactly know how to do this...


I should mention that I have never actually done a DOE study - this is my first one so if someone can proof my strategy I would feel a lot better.

 

[Coleman Donnelly]

Nothing?

 

[Ronen E]

Nothing?

Patience, my friend...

Most of us are volunteering here, and some actually choose NOT to surf the forums on weekends

I'm no expert on DOE but I'm sure there are plenty of good books on the subject.

Cheers,
Ronen.

 

[Coleman Donnelly]

Agreed,

I was actually concerned that it would not appear in the last 24hr thread that I assume most people would check in the morning.

 

[Ronen E]

Agreed,

I was actually concerned that it would not appear in the last 24hr thread that I assume most people would check in the morning.

Not to worry.

The moderators here do a pretty decent job in bringing those unanswered threads back to the top. Besides, under "search" in the top menu bar there's an "unanswered threads" option. I use it sometimes when particularly bored... I assume I'm not the only one.

Anyway, have a good one. Ronen.

 

[Stijloor]

Agreed,

I was actually concerned that it would not appear in the last 24hr thread that I assume most people would check in the morning.

Threads/posts not responded to will always get bumped the next day or following days.

Stijloor, Forum Moderator.

 

[harry]

Agreed,

I was actually concerned that it would not appear in the last 24hr thread that I assume most people would check in the morning.

Thank you. Actually, the best days to bump a thread is from Tuesday to Thursday when the traffic is at its peak and the best time is from morning (Cove time) onwards.

 

[Miner]

Center Points can be used to provide three different functions:

1. The first function is to identify whether curvature may exist in one or more of the factors.
2. The second function is to provide an estimate of pure error. This minimizes the need for many replicates of the experiment.
3. Finally, if distributed throughout the experiment, they provide evidence whether there may be a lurking factor that is changing in the background of the experiment.

Star (Axial) points may be added to an experiment AFTER curvature has been demonstrated to identify which factor(s) are responsible and to model that curvature using a quadratic equation. Normally, star points are not used unless significant curvature is indicated.

Power depends on your situation, but 0.70 is fairly typical. 0.90 would normally be considered quite high.

I do not recommend using Cp/Pp as your response variable. Use the mean and standard deviation directly as your responses.

Given your objective, you do not need replicates so much as repeats. Since you are not trying to exactly quantify the variation as much as to indicate directionality, you do not need 30 repeats. You should be able to use 8 - 10 repeats per run to get good results.

I recommend that you start small then perform follow up experiments as you learn more information.

 

[Coleman Donnelly]

Thank you very much for the response.

Agreed i will use sigma as my response, however i need to justify my decision to run 8-10 repeats with data that supports this as a requirement. How do i justify running 8 or 10 repeats and also demonstrate that this will be enough?

Thanks

 

[Miner]

The goal of early experiments is not to quantify the variation exactly for each experimental run. You simply want to identify whether the variation at one level is greater or lesser than another level.

Remember also that your experiment has 8 corner points. This means that for 8 - 10 repeats, you have 4x (8-10) or 32 - 40 values at each level. Again, this will not provide the final capability, but you will know which level has the least variation.