DOE - Qualitative Response - 4 Variables at 3 Levels Maximum

I am trying to conduct a DOE that will optimize the input variables so that I will receive a defect free product for a process.
This is actually my first DOE evensong I have done some course.

I am really struggling to get a start. I would really appreciate if someone can help me to get started.

I have say 4 variables at 3 levels max.
The output is just we the say the product is acceptable or rejected.


Any suggestions.
Thanks in advance for the help.

Sriram

Sriram,

Welcome to the Cove!

My first suggestion would be to see if there is a way to measure a continuous variable for the output, rather than just pass/fail.

For example, instead of a go/no-go gage for a dimension, measure the actual size. With a relatively small number of parts and variable data, you can spot fairly small differences. With pass/fail attribute data, you need a large sample - basically you have to find the percent defective and look for a change in the percent. Also, with pass/fail, you don't know if yu are too big or too small.

Tim

Thanks Tim for your reply.

I tried to get some form of measurement, but in vain.

My output is a product that we visibly check for defects. So i cannot measure anything from the output. we trying to optimze the factors so that we can reduce the rejection rate

Is there any way to set up the experiment and anlyze the factors?

do you have any example that I can refer to for qualitative responses

There are several standard DOE plans that will work with 4 factors and 3 levels. Box-Behnken and Central Composite designs could both work well.
Either of these would take about 30 trials (as opposed to 81 for a full factorial). Any good DOE book or software package should cover these designs.

If you have more than 4 factors, I would suggest an intial screening experiment to narrow down to 3-4 key factors, but with just 4 to begin with, you should be able to jump right to the designs listed above.

Also, even with visual inspection, there may be some way to quantify the defects - measure the length of scratches, count paint bubbles, estimate the degree of discoloration, etc. If you can do this, then it may give you a much more effective experiment and much more useful analysis.

Tim F

Tim Folkerts said:

Also, even with visual inspection, there may be some way to quantify the defects - measure the length of scratches, count paint bubbles, estimate the degree of discoloration, etc. If you can do this, then it may give you a much more effective experiment and much more useful analysis.

Tim F

Click to expand...

This is a good point, plus record the location and frequency.

I am unable to find examples for the design with the qualitative response.
can some one help me ?

every example i see is with numeric responses.
so still not able to get to the concept of qualitative responses.


Srira,

You need to differentiate a design from its analysis. The designs suggested would work as far number of conditions. Here are some thoughts and advice.

1. For each condition run about 300 parts and get the percent defective.
2. If parts are expensive, then look into logit analysis with binary responses.
3. Look into fractional factorial. Since you are not optmizing the response on a continuous scale, most likely any curvature you would see in the model would be random measurement error (visual inspection error).

Hope it helps some.

Re: DOE - Qualitative response - 4 variables at 3 levels max

Hi Sriram,
I'm also embarking on a series of DOE with discrete data to improve on the IC handler in my company.
The response for the DOE is a measure of the visual mechanical defects (discrete data), in particular, mold gap as found in the IC.
The mold gap on the IC is causing the IC copper trace to break off leading to open failures during testing.
We find it quite impoosible to convert the defects into variable data.
Hence we decided to stick to dicrete data as our response.
I've done up the design matrix for the DOE which in our case is a 6 factors, 2 levels, 16 run, Res IV fractional factorial design.
Pls find attached the pdf file on my DOE design matrix.
However, I'm still pending for the customer to release the handler for us to perform the experiments that would only happens on the last week of Sep 06. Therefore, the attached file will not contain any data and analysis at the moment. I'll post my DOE results up in the forum once completed.
Nonetheless, I found one excellent web site that provide an overview of DOE with discrete data.
Here's the link https://www.engr.wisc.edu
Pls scroll down to r119 of the website to download the material.You will need to provide some basic information to download the files.
To add on, I'll be using the following graphical methods to analyse the DOE data.
1. Normal probability plot of standardized effects
2. Residual plot vs fitted value
3. Resdual plot vs order
4. Residual plot vs variables
I would suggest that you keep your 1st DOE as simple as possible and if you only have 4 factors, I would recommend you to perform a full factorial on 4 factors and 2 levels instead.
Although not much but hope it helps.
thks.
jeffrey chang

Mr

jeffrey_Chang said:

Hi Sriram,
I'm also embarking on a series of DOE with discrete data to improve on the IC handler in my company. <snip>

Click to expand...

Jeffrey, that is an excellent reference. I want to bring to your attention that the sample size for each run condition is fairly large. If you do not have a large enough sample size, those transformations will not yield constant variance.

Hi Steven,
I do agree that the sample size has to be relatively large for discrete data study but then at times the situation simply don't allow us to have the luxury of a large sample size. Take my instance, I'm currently from the OSAT (Outsource Semiconductor Assembly & Test) industry and most of our equipments/ products are consigned from and owned by the customers. As such, to obtain a large enough sample size is quite difficult or practically impossible as approval is almost always required from the customers. As usual, they are skeptical in approving large sample size for such uses. Hence, our hands are always tied down due to such circumstances and we just have to make do with whatever we have on hand.


Thks & Rgds.
Jeffrey.