What can I treat as an input factor in Minitab?

[christianos]

Hi,

As I found this forum very useful I decided to ask experts again.

I was wondering if when using DOE (Taguchi approach) can I treat different type of coolants as an input factor. I mean, if a process is, for example, machining, I have input factors like: feed (feed1 feed2 feed3), speed (speed1 speed2 speed3) and type of coolant (coolant1 coolant2 coolant3). Doing ANOVA in Minitab, software can predict the best cutting conditions, however there are not always the best. Theoretically, speed 1, feed 1 are working the best. But these parameters were good for coolant1. But Minitab predict that it should be speed1 feed 1 and for example coolant3. For Taguchi, factors should change linearly, but definitely coolant is not like that.

Any suggestions how can I compare different types of coolant? What approach will be the best? Taguchi?

, christianos

 

[Miner]

ANOVA and DOE can be utilized for the following types of input factors with a continuous response variable:

• Continuous factors treated as discrete (e.g., speed at a high and a low level)
• Discrete factors (e.g., Supplier A versus Supplier B, Coolant A versus Coolant B)

The final analysis is affected by whether the factor is fixed or random, so read up on this aspect.

Judging from your comments, there must be an interaction between coolant type and one or more of the other factors. Therefore Taguchi methods are not appropriate. Taguchi orthogonal arrays are resolution III, which means main effects are aliased/confounded with 2-way interactions. You will probably want a resolution V design. If you have an understanding of which interactions may be important, you might get away with a resolution IV design.

Once you have selected the best coolant type, a response surface design would help optimize the remaining levels.

 

[christianos]

Ok, I'm confused now ;/

I attached some jpg files with plots from Minitab. Maybe this would help to understand my point of view.

I read a lot of papers where different people use Minitab for similar experiments (machining). The only difference was they did not use any type of coolant as a input factor. They use usually , feed, speed, type of material or point angle as a input factors.

According to my attached jpg files, could you tell me that Taguchi is not working definitely in my case?

In these statistical methods there is a very thin boundary between how to use each method for each case.

Thank you

 

[Miner]

This helps me understand some of what you are trying to accomplish.

It appears that you are using an L27 orthogonal array. As long as you placed your three factors in columns without aliased interactions, resolution will not be an issue. You can check this in the session window under the alias structure.

Regarding your question about coolant, coolant is what is called a fixed factor. This means that the levels that you investigated are the only levels of interest, and you are not using the results to apply to a wider population of coolants. It also means in this case that there is no directionality to the levels. That is, one of the three is the best, period. I would select the coolant that provides the best results, then optimize the other factors at the level of coolant.

You did not include your session window results, so I cannot determine the statistical significance of any of these. I am concerned about the number of missing data. Is this the same experiment as your other thread on missing data? I am also concerned about the appearance of your residuals vs. order graph. Did you randomize your runs? The missing data may also have contributed to this. Do all of the *'s signify missing data? That is the Minitab convention, but I want to verify that with you.

Finally, It does not appear that you ran an outer array or any repeats, so how are you calculating a S/N ratio? This approach only works with repeat data or outer arrays.

Note: if you attach your Minitab worksheet, some of us can assist in your analysis.

 

[christianos]

I tried to do 9 runs only (L9), however results were confused, so I thought would be better to do whole L27 to obtain more accurate results.

According to my results, I found that one of the coolants is significant and works better than others. In the near future I would like to use other types of coolants, I hoped that I might just add another coolant and compare with others from previous tests. Now I see it is not so simple.

I attached session windows for burrs and torque.

This is the same experiment as I described in my previous thread. Data are missing as in some cases (for some cutting conditions) a tool was broken and I could not get any reasonable results. All *'s are missing data.

I randomized my runs. I chose at random a sequence for a 9runs for coolant 1 and I repeat this sequence for coolant 2 and 3.

Each coolant was tested in following order: 5->4->1->2->8->6->3->7->9 in 9 runs scale.
I did not use any outer arrays. As I assume there is no significant noises. In fact I agree with you that Mean plot values are reverse of S/N ratio. I am not sure what does it means but for some factors S/N ratio gave me different results than mean values plot.


In the future,probably, it would be better to do just 9 runs and repeat them. As L27 is quite expensive (27 tools).




​

 

[Allattar]

With the design that you have there the residuals are going to look a little wierd.

Anything appearing at 0 is coming out exactly as the model suggests. Actually its a fair indicator that you do not have many replicatess in the model.

I am not certain Taguchi is the best approach here. If your looking for significant results or settings maybe a standard factorial, or even a face centred central composite design to fit for curvature would be a better approach.

Mind you I am not a fan of the Taguchi designs at all, and would rather gnaw my own arm off at the elbow before using one . (Just to put my personal feelings into context)

Also I am a bit concerned as you do not have an outer array, which is surely the main point of a Taguchi design to test with different environment settings. Without the outer array you ahve to be better using a Factorial or a Response surface. Although an open question here would be a comparison of the benefits of a Taguchi design over say a factorial when no outer array is used.

Anyway the ANOVA is just telling you about means, and not fitting a linear model here. You will only get a linear model in 2-level factorial designs.
Also with this design you can only really read the interactions between Cooling*Feed which isnt significant.

If you had run this as a 27 run General full factorial you could have tested every combination and found all 2 way interactions.

Sorry this is turning into a bit of a negative post against Taguchi designs here.

Was there only one run at each point? I would still think a Binary logistic regression here would be useful, but is going to be difficult with only one result each trial, it wasnt showing much to me. I could get coolant as almost significant, but without the other terms. Otherwise its too little to go on there.

If you use the calculator store the result in column 8, but type this into the equation.

if(Torque=miss(),"break","Good")

Its a quick way of marking broken tools in that column.

Becuase there are no outer arrays, and there are only 2 df in the error, any SNR is very difficult to find, its based on only a few points (hence those residuals all sat at zero).

Personally I would take the best means results from here, I am assuming highest is best. Coolant 1, F low F medium, S high, S low. Then try a 2 level factorial design. With a few replicates so hopefully we could then estimate chances of failure.

Its a tough question this one.

 

[christianos]

Thank you for your answer. These are not good news for me that Taguchi is not working in my case. I am not sure what to do now.
What should I do to adjust this L27 (or even L9) to make it working. I mean what about this outer array? What information I should put there?

I thought that Taguchi will be useful in my case.

Maybe I will give a better description of my tests.

I am doing some drilling tests. Each run = one drill, so for L27 I used 27 drills. My input factors are feed, speed and type of coolant. Each factor is on 3 Level. A difference between coolants is on lubrication and temperature of cooling.
While drilling I am measuring torque and thrust. After drilling I am measuring hole size and roughness of the holes. It give me several results.

I know that I should do each run at least twice, but I assumed one is enough to determine the difference.
How I should understand outer array. Is it array with the results of repeated runs?
The aim is to find out which type of coolant is working better, produce better quality of the hole. Speed and feed are factors which help to evaluate it. It would be great if there might be a possibility to predict the best cutting conditions (good hole quality, low torque and force, small roughness).
I am not certain as there is something different than statistical approach to do that.

What approach would be the best to evaluate possible usefulness of each coolant.

 

[Miner]

I took a quick look at your data.

If each * denotes a broken tool, the data in the two Excel files are contradictory. A simple data sort in Excel shows that you should run Speed at the High level to minimize breakage in the Torque file, but shows low speed in the Burrs file.

Number of broken tools:
______Speed
File___ L M H
Torque 6 8 0
Burrs__ 0 6 8

Am I misinterpreting something or could data have been incorrectly entered?

Just a note: A full Factorial experiment of 3^3 would have been 27 runs and you would have had all possible interactions vs. the L27 orthogonal array. I would only advise using the orthogonal array as a screening design at 2 levels. You can still apply Taguchi concepts such as robustness using classical DOE's. In fact, setting the inner and outer arrays up as a split-plot experiment (inner array = whole plot; outer array = sub plot) is more efficient in the number of runs. The S/N ratio is intended for use only with the inner and outer array design construction.

 

[Allattar]

Don't worry you can still get information out of most trials or tests. Although not always with the ability to prove statistical significance. We can still get hints of where to look.

For instance see the attached graph produced from the Torque data. This does assume that * means broken tool. You will have to let us know what * means for the Burr data.

What you have unfortunately discovered though is that different DOE tools are more efficient in different scenario's. Part of the art/science of using them is knowing which DOE is the most efficient for use in a specific role.

I have a strong suspicion that Taguchi get used as a buzzword for DOE and becuase of that is used for more DOE analysis than it should be.

Anyway from your results Coolant 1 appears, just visually with the bar chart to give less failures. Speed High appears good as well.

Problem is this isn't proved significantly. However if you take from this High and Low speed. Maybe Coolant 1, actually you could try the others but it may be simpler with just one or two coolants. And run a 2 level factorial with 2 or 3 factors, and replicate the design a few times to get an idea about chance of failure at each level you can build on what you have here.

 

[christianos]

I took a quick look at your data.

If each * denotes a broken tool, the data in the two Excel files are contradictory. A simple data sort in Excel shows that you should run Speed at the High level to minimize breakage in the Torque file, but shows low speed in the Burrs file.

Number of broken tools:
______Speed
File___ L M H
Torque 6 8 0
Burrs__ 0 6 8

Am I misinterpreting something or could data have been incorrectly entered?

Just a note: A full Factorial experiment of 3^3 would have been 27 runs and you would have had all possible interactions vs. the L27 orthogonal array. I would only advise using the orthogonal array as a screening design at 2 levels. You can still apply Taguchi concepts such as robustness using classical DOE's. In fact, setting the inner and outer arrays up as a split-plot experiment (inner array = whole plot; outer array = sub plot) is more efficient in the number of runs. The S/N ratio is intended for use only with the inner and outer array design construction.

Hi,

As I tried to modify names in inner arrays, Minitab probably has changed something. I did not see this error before I sent it. Each * denotes a broken tool. Good to hear that Taguchi is not useless. I need to think how should I make an outer array. If I do L9 and repeat each run twice, does it mean I could do S/N ratio plot?

I've attached a zip file, a I cannot attach a Minitab file directly. Inside there are whole results which I get in the original state. Maybe this time will be possible to do an analysis.

Thank you for your help.