DOE from pass production data

[#1]

How can I get document or clip, how to do to optimize factor for the best results from the pass data.

[#2]

If I'm interpreting your question correctly you are asking how to analyze a DOE using historical data.

We all wish to extract as much knowledge as possible out of historical data. I've found that in almost all cases though the historical data was not gathered in a way that facilitated the answering of any project related questions. I believe there is an applicable quote from Tukey regarding this aching desire. I do wish you're more lucky than me though.

You could try to apply multiple regression on that production data, but my fear is it will be failing certain assumptions.

It's unlikely you're Y is a function of a single X. I would instead suggest establishing an experimental objective and then gathering some data to run a sequential experimentation approach. First widen you're options with some structured brainstorming and direct process observation. Once you've gathered more factors besides the one you want to optimize run a fractional factorial to see if they're significant. Are you really sure this single factor tells the whole story? Near optimum the response surface can be complex consisting of multiple factors and possible interactions. 

After the fractional helps you identify what factors could be important then start to optimize by applying a steepest ascent approach. Use sequential experiments to hone in towards optimum and when you think you have level settings straddling the optimum run one more and include 5 center points to determine if there's any curvature and check for repeatability. Then add axial points to finish the optimization.

Of course all of this is hypothetical and really depends on your specific situation, the data will guide. As long as you keep it small and sequential you have the option to pivot when results don't match expectations. And since you're running small experiments, you don't waste resources along the way as you surely get surprised and must pivot.

There are many experimental strategies besides this one, but it's my favorite and tends to be relatively straight forward.

Goodluck

[#3]

Thanks you very much. Actually I used to join JMP webminar to show that there is one company have a low yield problem, it seem like a seasoning problem but they could not find root cause. 
They use there pass data and analysis like DOE method. Finally they can optimise there process to maximize yield. 
In my case, the process is long around 6 months then it hard to Make DOE experiment plan since it take a long time and precisely control. 
I plan to use production data with gathering factors and response and use the variation data to find our effect factory and optimize parameters. 

[#4]

If I were you I'd start with making process behavior charts of the production data. To first answer the question if the process is stable. You need to know what type of variation you're trying to solve, special cause or common cause. Sometimes there is no "root cause" and the problem that happened was just simply due to random chance. Common cause is the type of random and expected variation. Special cause is assignable to some root cause that you can name. The distinction is important because our actions would be different. Special cause is find and fix the root cause. Common cause means studying the system and fundamentally changing the process, the design or both to reduce the variation inherent in the system. It's harder work, but this is where DOE and critical thinking can really shine.

This sounds like a possible EVOP (Evolutionary Operation) scenario. We use this when it's difficult to break into production to run an experiment. You can find information online I'm sure, the technique was developed by George Box in the 1950's. The general idea was if we're running the process anyway to produce things, why not run it in a strategic way so that it not only produces things, but also knowledge about the process itself. EVOP is not a single experiment, but a way to run production to learn and optimize while it's running.

If you mean the process just takes a long time to execute and that's why DOE is hard. I'd counter that these methods were initially developed for agriculture. They could only run a single experiment all year if the response was yield of the harvest. So I'd argue if the process takes a long time to execute that's just more justification to run intentional balanced and orthogonal designs so you can learn as much as possible about the factors. And my other favorite answer when someone says this is going to take a long time "Well, we better start today then."

If you try multiple regression I still fear you'll have issues with multicollinearity. But maybe you'll get lucky and the important factors aren't correlated.