My company extrudes jackets over large cables and often the product diameter from one job to the next is different because all of the jobs are custom orders. Because of this we have to estimate what extruder RPM to use during startup based on data collect from other jobs. To do this I have been looking for a past job that is closest in size to the job I want to estimate, calculate the jacket annulus area for each, and then use the ratio of the two areas as a factor to multiply the past run data by to get the estimate for the new run. Sometimes my estimate is close and sometimes it is not.

To reduce the error in my estimate I wanted to look at all of my past data and see which past runs are reliable to estimate from and which are not. They way I am attempting to do this is as follows: I entered run data (extruder RPM vs line speed) from about 45 past runs into Excel. I then normalized each data set by dividing the RPM by the jacket annulus area (because it is different for each job) and the line speed. Doing this gave me a table of values that is basically independent of jacket size and line speed (I think). RPM vs line speed is basically linear for what we do, by the way. What I now want to do now is compare each population (run) to see which past runs do not fit in with the majority of the others, assuming the major of the past runs follow the same pattern.

Is an ANOVA the correct tool for this? Is my method and data manipulation accomplishing what I think it is? If an ANOVA is the correct tool, what should I be looking for in the results?

Thanks for the help.

Regression analysis of the annulus area using RPM and line speed as factors would be a better method.

Regression analysis is used where the input and output variables are continuously varied. ANOVA is used where the input variables, though they may be continuous, are constrained during the experiment to specific levels.

Regression can be used on historical data while ANOVA usually requires planned experimentation.

Thanks for the reply and the explanation about when to use an ANOVA.

I don't understand how regression analysis of the annulus area with RPM and line speed as factors would help. The annulus area for a given run is maintained constant throughout the run by varying RPM as line speed changes so RPM and line speed are not independent variables affecting the annulus area. Also, there are multiple RPM/Line speed data points for each run, which would I choose?

I have 40 odd sets of RPM vs. line speed data for extrusion runs all for different size products with correspondingly different jacket areas. So tomorrow if we get a new job I will go find data from a past product and try to estimate the startup RPM for the new product from a past job the is closest in size to the new job. As I mentioned above sometimes the estimate is good and sometimes it is not. I want to know why. One of my assumptions is that the recorded data was made when the extruder was producing a jacket at the specified nominal outer diameter (I don't have data to verify this either way). So maybe my bad estimates come from runs where we were making a product significantly deviating from our nominal OD target. Or perhaps there is another factor.

It has been a few years since I have taken a statistics class but I seem to remember there being a tool that will tell me which set(s) of data out a whole group of sets is statistically different from the whole group. I want to excluded these sets (or runs in my case) so that I don't use them for estimating purposes.

I am not sure any statistical analysis would benefit you here. :D If setting the RPM is not highly correlated with outside diameter, then you might be well served to examine the process more closely and see what additional factors affect the outside diameter.

By using regression, you can enter multiple variables and see how much of the variance is explained by the model. So assume you have four variables that theoretically can affect the diameter, and given past data, explain 90% of the variance in the RPM; then you have a pretty useful tool to start working with.

You indicated that you think deviating from nominal OD may be affecting the data. If it is outside the OD, then assign a 1 to those runs.

Also, I see nothing in here about temperature. Wouldn't temperature fluctuations possible affect the process?

Also, you state that you vary the RPM throughout the run. What compels someone in the process to vary the RPM? What are they observing?

Also, could there be differences with who sets it up, or if it is AM/PM?

I don't know that any of these will matter. I'm just thinking up some potential variables that might influence this.

You are correct that ANOVA will tell you if there is a difference, but if there is, what do you do then?:D Regression might better facilitate manipulating the variables to see what contributes to the different diameters.

I still recommend performing a regression analysis. Based on your latest feedback, I also recommend adding another factor.

Response = annulus area (or possibly outer diameter or wall thickness)
Factor A = actual wire diameter
Factor B = line speed
Factor C = RPM

The idea is that the response is what you are trying to control. It is affected by changes in the factors. Once you determine the equation relating the four, you can then input actual wire diameter, line speed, desired annulus/OD/wt and solve for the RPM.

Use 100% of your data in the regression analysis. Since you are adding wire diameter as a factor, there may be zero "outliers". Perform a check of the residuals to highlight outliers. Standardized residuals greater than 3 are suspected outliers.

If you organize your data in Excel and attach it, we can assist in the analysis.

In order for this analysis to be meaningful, the jacket material viscosity (or other lot-to-lot variation), wire diameter variation and temperature variable (if applicable) may also be important factors. You may find that lot-to-lot variation negates any correlation, or would be required input to the estimate.

My company extrudes jackets over large cables and often the product diameter from one job to the next is different because all of the jobs are custom orders. Because of this we have to estimate what extruder RPM to use during startup based on data collect from other jobs. To do this I have been looking for a past job that is closest in size to the job I want to estimate, calculate the jacket annulus area for each, and then use the ratio of the two areas as a factor to multiply the past run data by to get the estimate for the new run. Sometimes my estimate is close and sometimes it is not.

To reduce the error in my estimate I wanted to look at all of my past data and see which past runs are reliable to estimate from and which are not. They way I am attempting to do this is as follows: I entered run data (extruder RPM vs line speed) from about 45 past runs into Excel. I then normalized each data set by dividing the RPM by the jacket annulus area (because it is different for each job) and the line speed. Doing this gave me a table of values that is basically independent of jacket size and line speed (I think). RPM vs line speed is basically linear for what we do, by the way. What I now want to do now is compare each population (run) to see which past runs do not fit in with the majority of the others, assuming the major of the past runs follow the same pattern.

Is an ANOVA the correct tool for this? Is my method and data manipulation accomplishing what I think it is? If an ANOVA is the correct tool, what should I be looking for in the results?

Thanks for the help.

I'd use a control chart so that you may maintain the time sequence of the data.

I don't think that is a good idea,i cann't see anything releasing between them.