Taguchi's Analysis using Minitab 16

[Miner]

In Reply to Parent Post by Allattar

Just wondering about things here.
An L27 design for 4 factors and 3 levels on each factor. The standard design gives only 9 unique different combinations of those four factors. There are then 3 measurements at each of those 9, hence a 27 run design.

So 8 degrees of freedom would come from 9 - 1. Which would indicate that the linear analysis here is only analysing the means of those 9 combinations. Which makes sense if we assume that the Standard deviation is not constant across the design, as we assume that the variance is constant across a GLM.


Does that explain the missing df? as it isn't using 27 runs, but is just using means of the 9 different run combinations.

I reviewed Minitab's help documentation. I am not 100% positive, but I believe that Allattar is correct in his post. It would definitely explain the results. The only way to be certain, would be to manually calculate the ANOVA, or trick Minitab into it.

I was taught Taguchi methods by Shin Taguchi in the mid-80's. The use of the Orthogonal Array was taught separately from Robust Design, which analyzes means and S/N Ratio. This allowed you to think and decide how you would analyze the results of the DOE.

Minitab has highly automated the Taguchi analysis routine to the point that you cannot decide which analysis that you want to perform. It has made the decision for you and has not explained what it has done.

The only way to obtain the p-value is to pool the factor with the smallest MS, or perform a non-Taguchi analysis.

[privilegeastros]

Allatar has a good point on that! However, when im doing the prediction, the mean result is negative, where for a "small is better" the result data has to be a non negative value as can be to zero.

[privilegeastros]

in another two similar predictions, the values come negative , one for S/N ratio and to the other the MEAN value. what does the negative sign means?

[wang1568]

Hello privilegeastros
I found the problem as below
1) When the minitab to create a 3-leval 4 Factor with default Taguchi design, the default designs simply takes the first 4 of columns of the orthogonal array.
2) Watch the design on orthogonalTaguchi.MTW or minitab output (with step 1), it seem to have 27 runs but just 9 different combination, so is a L9 * 3(replicate) design, and its the reason you loss df=19 (got df=9 not DF=27)
3) The main cause is 4 factor are confounded when 4 factor in column 1~4, so the design is failed
4) The correct design is using column 1 2 3 and 5, not column 4

[sethu89]

Hello miner.
I am new to this concept. So help me with this design.
I need to choose a array from this.

3 factors : cutting speed, depth of cut, feed.

levels: cutting speed - 12,25.
depth of cut -10,20,30,40,50.
feed - 10,20,30,40,50.

what is the array to select?
and how to proceed with minitab??

Please help me..

Thanks in advance.

[Miner]

What is the objective of your experiment? Is it to determine which factors have a significant effect on the mean of the output, which factors have a significant effect on the variation of the output, or is it to determine the optimum levels at which to set the process/

[sethu89]

Thanks for the immediate reply.

With the above said values i am conducting experiments. The machining force and temperature will be measured.

I have to get the influence of each factor for the force and temperature.
And come to a conclusion that which is the best set of combination to have a better force and reduced temperature.

Thanks again.

[Miner]

I recommend that you take the following sequential experimentation approach:

1. Start with a 2-level full or half-fractional factorial design with center points.
   levels: cutting speed - 12,25.
   depth of cut -20,40.
   feed - 20,40.
   Center point = 18.5, 30, 30
2. If center points (curvature) are significant, augment the design by adding axial points to create a circumscribed Central Composite Design and additional center points to create a quadratic model.
3. If center points are not significant a linear model is sufficient to model.
4. Use Response Optimizer to determine the optimal settings for your two responses.

I recommend against using the Taguchi Orthogonal Arrays. Yes, they do work, and I personally have used them successfully, but have learned better approaches since then.