Hi
Could someone help me through a DOE please, its sometime since I used this tool and feel a little rusty.

I would like to use this tool to determine the critical quality characteristics on a heat seal machine. This would be the first part of the validation work that I intend to run.

Rather than run every combination of the variables I would like to identify the key variables that can influence the heat seal made.

The equipment forms a heat seal on 80µ polythene bags, this thickness of the material is a constant.

The three variables on the machine are
Feed speed
Seal temp
Pause / dwell time

the feed speed is expressed as a percentage - default percentage at present is 150%

The seal temp default is 165 C

The pause / dwell time is 1.3 seconds

Ideally I would like to run a DOE with the three variables to try and identify the key operational aspects.

Would like to run each variable at three levels , for example
Feed speed at 140%,150%,160%
Temp at 155,165,175
Dwell time 1.0,1.3,1.5 seconds

I have access to MiniTab, would greatly appreciate some help.

Many thanks

This sounds like a classic Response Surface experiment. The two most common approaches (both of which are supported by minitab) are Box-Behnken and Central Composite Designs. Either should work, but Box-Behnken might be easier.

Box-Behnken has only three levels, as you wanted. (CCD typically has 5 different levels which are not evenly spaced, although you CAN adjust a few things and get 3 levels). A BB design would use 12 points "around the edge" of the design, and typically 3 center points. A full factorial design would have 27 points (and 29 trials if you did three repeats at the center point). So you would be running fewer then 1/2 as many trials.



Tim

P.S. I might suggest using times of 1.1, 1.3, and 1.5 so that they are evenly spaced.

What is the measured output used to evaluate your variables?

Sorry forgot to mention this.

The bags will be tested to destruction on a dedicated bag testing machine. The machine inflates the bag to its burst strength.

The welded seal should be stronger than the bag material - hence I'll get a failure in the body of the bag material - not a the welded seam.

Therefore, I'm trying to define the weld settings at the boundaries of the process.

Too hot and I'll get the bag material melting and producing a poor visual seal.
Too cool, and I get lack of fusion between the seal, resulting in the bag opening under (little) or no pressure.

Test is measured in mBar.

Trying to find the optimum setting to allow for some process drift to create a robust process.

Sorry forgot to mention this.
The welded seal should be stronger than the bag material - hence I'll get a failure in the body of the bag material - not a the welded seam.


This is going to mess up your results. Your data will be a measurable, which will not be a function of the weld - always. The burst strength of the bag (not the weld) does not relate to the weld - and therefore the weld parameters. So, only the data for when the bag fails at the weld is valid to determine the weld variables.

Not sure I made that clear, but your measurable is a problem. It is great if you wanted to evaluate the blow molding process that makes the bag material. Not good for weld data.

But, don't feel bad, it usually is the toughest part of the DOE design. :cool:

This is going to mess up your results. Your data will be a measurable, which will not be a function of the weld - always. The burst strength of the bag (not the weld) does not relate to the weld - and therefore the weld parameters. So, only the data for when the bag fails at the weld is valid to determine the weld variables.


It's not a problem if you treat the data correctly. Your data can be treated like categorical data weld passes or fails, or you can treat it as truncated data - continuous when the failure is at the weld and pass or maximum when the weld passes adn the bag break.

cagtegorical data will effect your sample size.

I agree with Bob that the the response is very problematic.

If you ONLY are interested in WHICH fails, you could do a pass/fail test and count the percentage of time each fails. (This seems to be what Bev is suggesting.) I expect that you would get a rather sharp change - either the bag will fail most of the time or the seal will fail most of the time. This might make it hard to get a good fit and make good predictions about seal strength.

Also you say, for example, a high temp could degrade the bag strength. In this case, the seal itself could remain the same strength at the different temperatures, but the failure mode would change.

If you really want the seal strength, you should find a way to better measure that.


Tim F

Too hot and I'll get the bag material melting and producing a poor visual seal.
Too cool, and I get lack of fusion between the seal, resulting in the bag opening under (little) or no pressure.


Even pass/fail may make conclusions from the results very coarse, at best. You have two divergent criteria, one visual, one physical - each evaluated differently. That is complicates the resulting relationship. If you go to pass/fail, you leaped out of the variable measure to attribute. I think we all have a good idea what happens then.

I do not envy your dilemma. :(

But, coarse is better than nothing. :cool:

Just for kicks, did you ask the manufacturer of the machine how they qualify it? I have to think they have been down this road before.

Hate to reinvent the wheel.:cool:

There are several things to consider:

How does the seal normally fail? is the burst test representative of this or is a directly applied pull test horizontally or perpendicularly to the seal more representative? Sometimes I find that using worst case use condition test a better approach but this typically keeps you in categorical data. Remember the important thing here is to get a test that represents the real world not simply one that gives you the convenient data or sample sizes.

Certainly continuous data typically provides richer information, but don't despair if you are stuck in categorical pass/fail results...it may be more 'course' but sometimes this is the world we live in and you can gain adequate knowledge of your process with categorical data. You will need to use larger sample sizes and it will be easier to use confidence intervals for the ratios instead of standard hypothesis tests but this is doable.

EXACT confidence intervals for Binomial data (needed for small failure rates and zero failures in the sample):

Lower Confidence Limit: BETAINV(α/2, d+1, n-d+1)

Upper Confidence Limit: BETAINV(1-α/2, d+1, n-d+1)

where α = alpha risk, d = number of defects, n = sample size

I have solved many complex rare defect rate Problems that were categorical in nature and impossible to translate to continuous data. I have also optimized input factors to ensure yield for categorical data. It can be done quite successfully, you just have to think about it a little more.


Another point is that you do need to be careful of limiting your factors to available "knobs". Examples here are that this type of weld is also influenced by pressurewhich may be a setting AND a function of the geometry of the fixture. so if you have mulitple cavities the different cavities can be different levels of the pressure factor. they may aslo have different temperature profiles so at a minimum you should test across cavities to capture this varation even tho you can't twirl a knob. The same concept applies to the material thickness - it does vary, nothing is ever constant. it can vary within a lot and between lots so I'd include multiple lots in the study... these kinds of "passive" factors will influence the output and even if they are not primary factors you need to understand that the settigns you choose for the "active" knobs' will perform across the normal variation of the passive factors...

There are several things to consider:

How does the seal normally fail? is the burst test representative of this or is a directly applied pull test horizontally or perpendicularly to the seal more representative?

Bev brings up a good point. I might add if the failure is happening at bag rather than the seal weld, maybe the thing to do is go so low (or high) on you setpoints to capture where the seal actually does fail (rather than the bag). If you find this area, you may have a better chance of knowing the process limits, the effect of each parameter on the weld, even though you would never normally produce a weld at those parameters. Two point tensile testing, three point tensile testing, burst, use what makes sense. Tensile testing can be tricky, because the bag material may show deformation (necking), which will look bad, but will not represent weld failure. If you can get it to work, it will answer the question which parameters have the most effect, and then you can focus on them.:cool:

Can you measure the amount of meltdown as a continuous variable? For example, a DOE for a welding process might measure the weld penetration depth instead of weld strength. This would provide more actionable results than the proposed measure.

Hi everyone, thanks for the feedback. I'll try and flesh out the process as there are alot of differing queries in response to my original thread.

I have polythene bags which vary in thickness, typically from 50 micron through to 125 micron.

The heat seal machine can be set to a temperature to create a heat seal weld, the seal is then "cooled" until a set temp is reached. For example 90c for the heat seal and the machine releases the heat seal bar at 85c.

The test equipment can take two forms of test.

1. Burst strength
The sealed bag has a washer applied to the surface of the bag and a stylus (or needle) pierces the bag. Air pressure is forced through the needle to inflate the bag to burst point. If the heat seal parameters are inadequate, the welded seal will fail. This failure usually takes the form of lack of fusion between the two webs.

If the weld does not fail, then the figure (max pressure at burst) is a measure of the material as opposed to the weld strength.

2.Integrity test
The same process with the needle or stylus is used, however, this time the bag is inflated to a known (predetermined ) pressure. This pressure is then monitored over a 5 minute period to check for a pressure decay. Pressure decay would indicate a hole in the bag. Again, this could be viewed as measuring the physical properties of the bag , as opposed to measuring the weld seal. It would be possible to use dye penetrant to determine where the hole was (if one existed).

I understand the issues put forward, but apart from determining the settings by visual assessment I'm slightly at a loss. I'd like to apply some form of variable approach to determining the settings.

Currently there is no assessment of any sort taking place, as the "visual" checks are not performed.

I'll have to have a re think on this - but could I not determine the boundaries of the weld performance in defining the upper and lower settings where a poor seal is made? Perhaps this no longer fits the DOE format though, and is a simple go / no go test?

DOE's do not require continuous data but in this case you also don't have to move to all categorical data either.

my advice is to run your original experiment. Don't use statistical software at first just plot your results. DON't use any data for bag failures - when you switch from a weld to a bag failure just use the highest seen weld failure at first.

the one drawback to this technique is that you will not know if you are just safe for weld strength or well safe, but at this point the important thing is to run the experiment. Once you are done - post your data here and we'll help you thru the more elegent analysis...

sometimes we get way too caught up in an elegent experiment that makes full use of standard statistical analyses in our canned stats sortware. The important thing is to learn somethign from the experiment - and you will do that.

Just for kicks, did you ask the manufacturer of the machine how they qualify it? I have to think they have been down this road before.

Hate to reinvent the wheel.:cool:

There are a number of interested parties in this validation.

Firstly, the raw material suppliers of the polythene bags. Rightly or wrongly, the raw material suppliers are unaware of the critical quality characteristic. The main purpose of these bags is to maintain product integrity. The end user product is terminally irradiated (gamma irradiation).

The equipment used to make the seals is supplied by different companies - therefore, I'm treating each individual machine as a seperate validation exercise. The equipment is used widely across the industry , so at least there is not too much confusion on this front.

The different operators using the equipment and raw materials. At the moment there are different people doing different things, as there are no operating parameters based on facts.

My initial aim will be to run some off line trials to try and understand the parameters, and hopefully come to some conclusions. These conclusions will need to be put into place in the production environment and see if the off line results tie up with the use in production on a routine basis.

The test equipmet.
I've contacted the manufacturers of the test equipment, but they haven't been too helpful. I've been largely left to my own devices here. I agree with some of the previous responses, the key to determine the weld characteristic(s) is a better measurement of the output. Ideally a tensile test or peel test would be more appropriate.
I've been told that I have to run the tests with the current test equipment though.

Hoping to get some initial work done soon.

Whether plotting "data" or completing a factorial matrix, it looks to me that you really have to boil it down to attribute level: did the weld fail (physically or visually)? The variable data can not be fully attributed to the weld. Even a tensile test data can intermingle bag failure and weld failure. It has a good chance of confounding the data and its interpretation. :cool:

on the other hand - my advice would be what it always is: stop admiring the problem and get some data. if we waited for perfect data we'd still be living in caves. :nope:

Knowledge and the scientific process is an iterative process...your first exploratory experiments can be sloppy - lacking in full balance and structure - as long as you learn something! Then you can design an experiment that makes sense for your situation.

I will make one last appeal concerning the search for continuous data: in theory it's the best thing. The grand majority of articles and text books deal with it, and even more irritating they stick with the Normal distribution to exclusion of almost everything else. well as Bob is so often wont to remind us, the world is not perfect. Stop looking for a 'great measurement system' and get one that is representative of how your bags fail in practice. If you can measure the bags in a way that is representative with field failures using continous data, then great. If you find a way to test the seal that gives you continuous data in a Normal distribution - but it isn't representive of how your bags fail - then your analysis in any given software package will be easy, elegant and completely wrong.

We have a very similar process in my current business; down to the use of bags from suppliers who have no clue what we use their plastic for and we irradiate our product at the end. We used to perform a burst test and a peal test. everything always passed with flying colors. the "bags" failed miserably. The cause: micro areas of non seal thru which the liquid leaked. we could only test it visually. but hey - we solved the Problem in the field using visula categorical data because that's we had. We determined the cause of the non seal and now we monitor it ("it" is measured using continuous data...)

Yep, Bev D makes alot of sense.

By posting this request for help,it has helped me understand the requirements and mull over the issues.

I have now produced a test protocol which is being circulated internally for approval.

I'll gather the data and draw some conclusions pretty soon.

As Bev D says - gathering this first set of data (even if its slightly wanting for finite detail) is the first step in establishing something that can be improved upon in the future.

By way of an update, I've found some interesting results.

I won't quote the actual figures, but give an overview of the results.

I used an 80µ LDPE bag as the raw material and a set of parameters for the heat seal machine.

I've determined the heat seal settings, by finding out the workable range of settings through a little trial an error.

The bag burst figures achieved are a reflection of the bag material itself - it has nothing to do with the heat seal settings used. The heat seal formed will fail if the settings are too low.

There are three critical parameters ,heat setting , dwell time, speed of the conveyor (heat seal head).
The most important parameter is (no real surprise) is seal temperature.

I suppose the real purpose of this exercise will be to provide documented evidence of the heat seal settings used, and to establish a SOP for the manufacturing guys.

I will then prove the validity of the settings in the PQ (by running 3 seperate batches).

Not so much a DOE at present, more like providing the basis for a future improvement and DOE.