What if DOE (Design of Experiments) is not possible?

pagnonig

Quite Involved in Discussions
#1
Dear all,

I'm approaching the world of Six Sigma and during my first readings I came up with a doubt.


I understood that I can use DOE in order to find the y=f(x) function for multiple x. Different techniques are available depending on the number of x and other parameters.

My question is... what happens if I can't perform DOE. How can I determine the most influencing Xs and consequently the f(Xs) function in case of multiple inputs?

I hope I'm not too confused...
Thank you.

Giuseppe:bigwave::thanx:
 

Duke Okes

Quite Involved in Discussions
#2
It depends on what you mean by "can't perform DOEs." Are you talking about processes you can't measure, or lack of process time needed to perform the experiment, etc. ...?
 

pagnonig

Quite Involved in Discussions
#3
It depends on what you mean by "can't perform DOEs." Are you talking about processes you can't measure, or lack of process time needed to perform the experiment, etc. ...?
Thank you for your answer.

It means that I CAN measure the process (e.g I can sample it) but I can't perform a DOE because of a lack of process time.
 

Duke Okes

Quite Involved in Discussions
#4
I CAN measure the process (e.g I can sample it) but I can't perform a DOE because of a lack of process time.
Why is there a lack of process time? Is there inadequate capacity to meet current requirements? Is it a 24-hour operation? Is the problem not serious enough to justify taking the time for the experiment? Other reasons?
 

pagnonig

Quite Involved in Discussions
#5
Why is there a lack of process time? Is there inadequate capacity to meet current requirements? Is it a 24-hour operation? Is the problem not serious enough to justify taking the time for the experiment? Other reasons?
Dear Duke,

It not a practical case, I just wonder which are the alternatives to a DOE.
 

BradM

Administrator
Staff member
Administrator
#6
If I understand you correctly, you are approaching this as a Regression Analysis. Namely, condition Y as a function of X (or multiple X's).

So, you are producing widgets (Y). You are wanting to identify the different variables (X) that will affect (Y) during production. However, you cannot manipulate any of the X values while it is running. Is that correct?

First, approach it from a common sense perspective. If you are making widget, on the surface what factors seem reasonable that would affect it? This is important. You cannot have a bunch of variables that appear with numbers to be related, but have no association. Your variables need to make sense.

Begin measuring these variables, and the states of your Y (output). With even data, you can develop a regression equation (or whatever type of analysis you would like).

There is more than one way to do research. Experimental design is where you manipulate one variable to observe the effect on another. However, a lot of research does not allow variables to be manipulated. Therefore, a theoretical case is made for a relationship, measurements are made, and the case is supported or not supported.
 

Steve Prevette

Deming Disciple
Staff member
Super Moderator
#7
I understood that I can use DOE in order to find the y=f(x) function for multiple x. Different techniques are available depending on the number of x and other parameters.

My question is... what happens if I can't perform DOE. How can I determine the most influencing Xs and consequently the f(Xs) function in case of multiple inputs?
This is precisely the advantage of SPC charts on an ongoing basis. Most processes have natural fluctuations in the input variables on a routine basis. You can plot the output variable on a control chart.

If there are significant trends (up or down) on the SPC chart, you can examine what the input variables were doing at the time of the trend and may be able to infer the y = f(x) relationship.

If the output variable is stable, then the assumption is made that the routine variation in the inputs doesn't have an effect on the output. Now, if the output is at a "good" level, you are in good shape. If it is stable in need of improvement, it does get a little tougher.
 

pagnonig

Quite Involved in Discussions
#8
If I understand you correctly, you are approaching this as a Regression Analysis. Namely, condition Y as a function of X (or multiple X's).

So, you are producing widgets (Y). You are wanting to identify the different variables (X) that will affect (Y) during production. However, you cannot manipulate any of the X values while it is running. Is that correct?

First, approach it from a common sense perspective. If you are making widget, on the surface what factors seem reasonable that would affect it? This is important. You cannot have a bunch of variables that appear with numbers to be related, but have no association. Your variables need to make sense.

Begin measuring these variables, and the states of your Y (output). With even data, you can develop a regression equation (or whatever type of analysis you would like).

There is more than one way to do research. Experimental design is where you manipulate one variable to observe the effect on another. However, a lot of research does not allow variables to be manipulated. Therefore, a theoretical case is made for a relationship, measurements are made, and the case is supported or not supported.
Brad, thank you for the excellent answer.
So I understand that if I can't carry out DOE's I have to apply a regression analysis in order to obtain the f(X) and to understand how the measured Xs influence the Y.
Any suggestion about docs or web material to read on this subject?
 

pagnonig

Quite Involved in Discussions
#9
This is precisely the advantage of SPC charts on an ongoing basis. Most processes have natural fluctuations in the input variables on a routine basis. You can plot the output variable on a control chart.

If there are significant trends (up or down) on the SPC chart, you can examine what the input variables were doing at the time of the trend and may be able to infer the y = f(x) relationship.

If the output variable is stable, then the assumption is made that the routine variation in the inputs doesn't have an effect on the output. Now, if the output is at a "good" level, you are in good shape. If it is stable in need of improvement, it does get a little tougher.
Thank you Steve, excellent answer as well!
 

Duke Okes

Quite Involved in Discussions
#10
Without knowing why you can't do the DOE it is difficult to know what alternatives might fit to your situation. However, here are some that might:

- use EVOP (more likely if chemical-type and/or continuous process)
- use component swapping (more likely if mechanical/electronic product) or other Shainin techniques
 
Top