We are running a DOE to determine the best shape of a speed vs. time curve. Our moulding machine allows us to split the injection speed into 10 segments to allow controlled flow of plastic into the part. We know that at the end of the injection the speed should be lower.

Seg 1 - 7 Levels are 2.2 - 3.4

Seg 8 levels are 1.8 - 3.0

Seg 9 levels are 1.0 - 2.4

Seg 10 levels are 0.5 - 1.1

Restrictions:

Segment 1 - 7 are greater than or equal to 8 and 9

Segment 8 is greater than or equal to 9

Segment 10 is less than 1-9

DOE (Response Surface 4 factors, 2 levels, 2 blocks)

The items marked x in the restricted columns I would not be able to run due to restrictions.

RunOrder SEG1-7 SEG8 SEG9 SEG10 Restricted.
1 3.4 3 2.4 1.1
2 2.2 1.8 1 0.5
3 2.2 3 1 1.1 x
4 2.8 2.4 1.7 0.8
5 2.2 1.8 2.4 0.5 x
6 2.2 3 2.4 1.1 x
7 2.2 3 2.4 0.5 x
8 3.4 3 2.4 0.5
9 3.4 3 1 1.1 x
10 3.4 1.8 1 0.5
11 3.4 3 1 0.5
12 2.8 2.4 1.7 0.8
13 2.2 3 1 0.5 x
14 2.2 1.8 1 1.1 x
15 2.8 2.4 1.7 0.8
16 2.8 2.4 1.7 0.8
17 3.4 1.8 1 1.1 x
18 3.4 1.8 2.4 1.1 x
19 2.2 1.8 2.4 1.1 x
20 3.4 1.8 2.4 0.5 x
21 2.8 2.4 1.7 0.8
22 2.8 1.2 1.7 0.8
23 4 2.4 1.7 0.8
24 2.8 2.4 1.7 1.4
25 2.8 2.4 3.1 0.8 x
26 2.8 2.4 1.7 0.8
27 2.8 2.4 1.7 0.2
28 2.8 3.6 1.7 0.8 x
29 1.6 2.4 1.7 0.8 x
30 2.8 2.4 0.3 0.8 x

Any suggestions??? Do I ignore the restricted items? Do I run them with different values? Any help would be greatly appreciated.

Cheers,

Jen

Jen,

This looks like a CCD design, which is a good place to start. Unfortunately, the limitations that you have cut way back on the allowed combinations.

Of the 25 different sets of levels in the design, only 10 seem to be left after the restrictions. I'd be really wary about that few combination, especially since you would have 14 parameters if you are doing a full quadratic fit. You would end up with significant aliasing.

You could try adjusting some of the restricted runs to some close-by allowed set. That would be an improvement over just skipping the runs.

One "fancy" option would be a d-optimal design. You would need to find a list of some of the allowed combinations (I think there are 130 total in your case). Then use some software (like minitab) to choose an optimal subset to acheive the goals you desire. It takes a little more computer power to create the design and a little more to analyze, but it should give the most robust results.


Tim F

We are running a DOE to determine the best shape of a speed vs. time curve. Our moulding machine allows us to split the injection speed into 10 segments to allow controlled flow of plastic into the part. We know that at the end of the injection the speed should be lower.

Seg 1 - 7 Levels are 2.2 - 3.4

Seg 8 levels are 1.8 - 3.0

Seg 9 levels are 1.0 - 2.4

Seg 10 levels are 0.5 - 1.1

Restrictions:

Segment 1 - 7 are greater than or equal to 8 and 9

Segment 8 is greater than or equal to 9

Segment 10 is less than 1-9

DOE (Response Surface 4 factors, 2 levels, 2 blocks)

Any suggestions??? Do I ignore the restricted items? Do I run them with different values? Any help would be greatly appreciated.

Cheers,

Jen

I have a few options/suggestions below. Do not ignore the restricted items. Big Mistake.

Option 1: Use the d-optimal approach that Tim Folkerts suggested. However, this is complicated and should be approached with care. See http://www.itl.nist.gov/div898/handbook/pri/section5/pri521.htm

Option 2: Use the concept of Sliding Levels. This is a Taguchi approach but, used with care, does work. See http://www.gaasmantech.org/Digests/2002/PDF/11e.pdf and http://files.aws.org/wj/supplement/05-2002-ALLEN-s.pdf for examples. This requires some prior understanding of the process to establish sliding levels. An example of this is curing rubber. Rubber cures faster at higher temperatures, so the cure time must slide with the temperature. At 350 degrees, the time may be 90 and 120 seconds, while at 450 degrees, the times may be 30 and 45 seconds.

Option 3: Use EVOP or Evolutionary Operation. See http://www.visteon.com/utils/whitepapers/20_ccwdq.pdf for a White Paper on this approach. This is fairly simple and requires no prior knowledge of the process. If your main interest is optimization without developing a model of the process, this is your best option.

I hadn't seen the idea of sliding scales before. It looks like an interesting option, although deciding how to set the levels would be challenging, and analysis doesn't seem like it would fit standard schemes, so you might have trouble finding software ready-made to help.

I had forgotten about EVOP, but it could be a good approach. If you are in production already and it is difficult to stop production to do experiments, EVOP slowly adjusts values to find an optimum setting. It may not get to the best spot quite as fast, but it also allows production to keep running more smoothly.

Good suggestions, Miner.


Tim F

P.S. None of these approaches are the typical beginner DOE's. Use them with caution - perhaps find a mentor/consultant/buddy to help if you don't feel comfortable.

P.P.S. Here's a new idea I'm making up as I go - kind of along the lines of Miner's sliding scale idea. Keep the levels you have for the first factor. For the other factors, have them how much you decrease the settings from the previous level. For example
Seg 1-7: 2.2 - 3.4
Seg 8: 0, 0.1, 0.2, 0.3 or 0.4 lower than Seg 1-7
Seg 9: 0, 0.2, 0.4, 0.6. 0.8 lower than Seg 8
Seg 10: 0, 0.2, 0.4, 0.6. 0.8 lower than Seg 9I'd have to think if there is any major challenge in the analysis, but off-hand it seems it should work with standard analysis approaches. You would have to think if this range of variables makes sense for your process and perhaps adjust accordingly.

I hadn't seen the idea of sliding scales before. It looks like an interesting option, although deciding how to set the levels would be challenging, and analysis doesn't seem like it would fit standard schemes, so you might have trouble finding software ready-made to help.

I had forgotten about EVOP, but it could be a good approach. If you are in production already and it is difficult to stop production to do experiments, EVOP slowly adjusts values to find an optimum setting. It may not get to the best spot quite as fast, but it also allows production to keep running more smoothly.

Good suggestions, Miner.


Tim F

P.S. None of these approaches are the typical beginner DOE's. Use them with caution - perhaps find a mentor/consultant/buddy to help if you don't feel comfortable.

P.P.S. Here's a new idea I'm making up as I go - kind of along the lines of Miner's sliding scale idea. Keep the levels you have for the first factor. For the other factors, have them how much you decrease the settings from the previous level. For example
Seg 1-7: 2.2 - 3.4
Seg 8: 0, 0.1, 0.2, 0.3 or 0.4 lower than Seg 1-7
Seg 9: 0, 0.2, 0.4, 0.6. 0.8 lower than Seg 8
Seg 10: 0, 0.2, 0.4, 0.6. 0.8 lower than Seg 9I'd have to think if there is any major challenge in the analysis, but off-hand it seems it should work with standard analysis approaches. You would have to think if this range of variables makes sense for your process and perhaps adjust accordingly.

Your suggestion is right on target for sliding scales. One big difference in the analysis is that: 1) we are assuming an interaction between these segments and compensating for it; 2) therefore the sliding scales, if done exactly right, will show a non-significant interaction between these factors, whereas a classical approach would show a significant interaction; 3) You have to use coded levels (e.g., 1, 2, 3) instead of actual levels to get an analysis; 4) You may want to regraph main effects using actual levels to better visualize the effects. This will overcome the lack of interaction significance.

Thanks everyone... I think we can change the last couple segments of the profile to a slope instead of a value. That will allow us to keep everything I have minitab design. Does this make sense? looking at it it definitely satisfies all of my restrictions.

There definitely is an interaction with these segments as it is an injection fill profile over 10 secs divided by 10.

Cheers,

jen