iSixSigma

Controlling for an Underlying Variable

Six Sigma – iSixSigma Forums General Forums Methodology Controlling for an Underlying Variable

Viewing 19 posts - 1 through 19 (of 19 total)
  • Author
    Posts
  • #54403

    roarty
    Participant

    To apply an epoxy coating to an off-shore steel structure, several washes with de-ionised water is required to remove salt contamination from the steel surface. Due to the cost associated with transporting large volumes of de-ionised water, I have been asked to develop an epoxy coating which can be applied direct to salt contaminated steel.

    I have 8 experimental coating formulations applied to 8 steel panels which are deliberately contaminated with salt. The performance of each coating is assessed by the number of days before failure in a laboratory test.

    My problem is that the level of salt contamination (in mg/m2) on each panel varies and I would like to control for this variable so that I can pick a winning formulation based on the merits of the coating and not because that particular formulation happened to be applied to a lightly contaminated panel.

    I am prepared to repeat this test schedule several times; each time applying the same 8 experimental formulations to newly contaminated steel panels which will of course vary in the level of contamination. One approach I am considering is to measure performance by dividing the number of days to failure by the level of salt contamination but this is undesirable as it assumes a linear relationship between contamination and failure period.

    I think there should be a statistical method for controlling for an underlying variable while testing for differences between experiments. I would appreciate any suggestions on an appropriate experimental design. Thanks.

    0
    #195228

    MBBinWI
    Participant

    @cognition – you will need to either measure the salt contamination, control the salt contamination, or increase your repetitions to normalize out the variation of the variation.

    0
    #195229

    MBBinWI
    Participant

    that’s supposed to read “normalize out the variation of the contamination.”

    0
    #195231

    Chris Seider
    Participant

    If the salt contamination is consistent on each sample, can you apply part of all 8 formulations to each sample?

    0
    #195240

    MBBinWI
    Participant

    @cseider – my understanding was that the amount varied across each panel, which was unknown. If it is uniform, then yes, painting each panel with all 8 formulations would be appropriate.

    0
    #195242

    roarty
    Participant

    Thanks for your help,

    I should have said that the variation in salt contamination is between panels not within the panels. In other words a single panel tested on each of the 4 corners will give pretty consistent readings but the average salt contamination reading for one panel will differ significantly from the next panel.

    The stainless steel panels are soaked in a saline solution for a period of time to achieve contamination so I can control concentration of the saline solution as well as soak duration, I can also rinse a highly contaminated panel with de-ionised water. Even with these measures I will struggle to get contamination readings on my panels which are within 10% of each other. I believe there may be a statistical method to allow for the variation.

    The idea to apply all 8 formulations to each panel is also not an option; the test method requires a single coating applied over the whole of the panel.

    0
    #195262

    Robert Butler
    Participant

    You said you have “8 steel panels which are deliberately contaminated with salt.”

    And you also said ” My problem is that the level of salt contamination (in mg/m2) on each panel varies and I would like to control for this variable so that I can pick a winning formulation based on the merits of the coating and not because that particular formulation happened to be applied to a lightly contaminated panel.”

    In other words – there is something amiss with your salt application process.

    If salt is coating out on the various stainless steel plates how are you maintaining the level of salinity in the bath? How are you insuring that the plates are seeing the same amount of immersion and the same level and type of fluid circulation across their surfaces? If there is some kind of fixture you are using for the simultaneous dipping of these plates how are you guaranteeing the plates are all seeing the same conditions in the bath?

    If you could provide a few more details about the coating process perhaps I could offer a few suggestions with respect to testing for and possibly minimizing the sheet-to-sheet salt coating differences.

    0
    #195263

    Yogesh
    Participant

    @cognition the information provided by you looks insufficient to provide specific solution. Here are few ideas, you may try.

    Use 8 formulations on single panel – if testing requires single coating as u said then may be cut each panel into 8 and then apply coating to all 8

    I somehow believe if you use a standard process, variation in contamination should be controlled, however if you can’t than try contaminating multiple panels ( say 100 numbers as quantity is not a problem for you) than pickup group of 8 samples which are contaminated in +- 10% range of each other. you may have to increase the sample size if need be.

    Even if nothing works you have contamination readings of each panel – apply each formula on multiple panels and establish corelation . Formula should look something like Salinity removed = ( Salinity before applying sample – Salinity post applying sample )

    0
    #195264

    I have nothing further to add other than some info. I found in the usual way:

    http://www.duckworksmagazine.com/04/s/articles/underwater/

    0
    #195266

    MBBinWI
    Participant

    @cognition – I would suggest either making the individual panels larger so that you can put a sufficient amount of coating on each panel, or (if I understand correctly) you could put multiple panels in the saline solution for the same amount of time, which should produce a similar corrosion level. Then you can paint each one with a different formula.

    0
    #195269

    roarty
    Participant

    My thanks to everyone who has taken the time to offer advice…

    Robert,

    You raise a valid point, it is essential that I subject each panel to the same contamination procedure. I must admit however that I have just placed all eight panels into a bucket of salt solution (5% by weight). As the water has evaporated I have added further additions of salt water to ensure the panels remain fully submerged – so the concentration of the salt solution has changed with evaporation over time. Also the panels are stacked upright and against each other in the bucket and this may also introduce further variation. So I am guilty of a haphazard approach which I shall take steps to rectify when I repeat this work.

    Nevertheless I have gone ahead and put all eight coated panels on the test cells. I had hoped that statistics would rescue something meaningful from my sloppy approach. For example, one of the eight coatings is a generic epoxy coating. I do not expect it to have any salt tolerant property but included it as a control. I thought I could test the control coating over several different panels of differing contamination levels then fit a regression line to the data for performance versus contamination level for this control coating. I can then draw the 95% upper confidence limit for the regression line then plot the data points for each of my 7 experimental coatings. The data points which lie above the 95% upper confidence limit indicate that experimental formulation is significantly better than the control. Do you think this approact has merit? What if the regression line is non-linear? Will this mean the residuals for the 7 experimental coatings is non-normal?

    Yogesh,

    Unfortunately cutting the panel into 8 is also not an option; the panel must fit the test cell. However, I like your lateral thinking and if I still cannot achieve consistent contamination levels with my improved contamination procedure then I may just have to follow your advice and contaminate a large number of panels and pick a subset.

    Andy U,

    The article you link to is typical of my application. I am brush applying my coatings to panels in the safety of my lab but someone much braver than me will have to paint steel structures out in the middle of the North Sea.

    MBBinWI,

    I agree, the panels should be exposed in the same saline solution for the same length of time. I need to control my method of contamination for the next time I do this.

    0
    #195272

    Robert Butler
    Participant

    So you have 8 experimental coating which,given your description of the coating matrix, would suggest you ran an 8 point experimental design with X number of variables in each coating.

    Based on the article cited by Andy U, it would appear that the last thing you want is uniformity with respect to salt concentration – what you really want to know is which combination of X variables provides the best “stick time” regardless of salt concentration. In other words – if the above is a correct summation of your situation – what you have is a classic experimental design problem. Therefore, number the sheets (take a salt concentration reading on each one if you want) randomize the coatings to the sheet and – presto – the lurking variable gets shunted over to the noise term in your regression model. Build the model, run some predictions, confirm the model, and make your adjustments accordingly.

    0
    #195280

    roarty
    Participant

    My 8 formulations are essentially the same formulation but with a different additive in each (the control does not contain an additive). To frame my design in DOE terminology; it is best described as one factor with 8 levels. So you can see it is not really a factorial experiment, it is more a traditional experiment to screen a number of different materials.

    You are correct however when you say that I am more interested in performance regardless of salt contamination level. Uniformity cannot be guaranteed when applied ‘in the field’. My desire for uniformity in the lab was only to ensure I was making like for like comparisons.

    So if I understand you correctly, you are suggesting several repetitions of my 8 experiments. The more repetitions I do then the more I can ‘average out’ the lurking variable that is salt contamination. This approach is possible although I had hoped there would be an elegant statistical approach which would cut down on the number of experiments while providing the same level of detail. In an ideal world I would find out; not only which of my formulations is best but also its level of performance versus salt contamination.

    Thank you for your help.

    0
    #195281

    I’m a bit out of practice these days, but one thing I’d consider is dividing each panel in eight portions and then painting each portion while the panel is submerged in a bath of salt water (proper grey sea-salt). I’d then adjust the salt concentration while painting each portion. Obviously, you could also dilute the bath by adding some fresh water, so with a little effort, you could randomize the portions, Whew! If you have eight panels you could repeat this process eight times. Perhaps I’m missing something …

    0
    #195282

    Chris Seider
    Participant

    I should have asked earlier but how exactly are you measuring success? Can it pass an MSA?

    0
    #195283

    @cseider Chris, I’ve assumed the response is uniformity of coverage under various concentrations of sea-salt. By uniformity I assume the main objective is to get a coat of paint on the panel. I’m afraid I can’t think of an application where structure can’t be pre-coated, but I’ve often been taken by surprise in the past. (One of the strangest processes I ever saw was spinning silk disks on carbon powder to try and produce diamond coatings!!!)

    0
    #195290

    Robert Butler
    Participant

    Maybe we need to talk about those 8 additives a bit more. As described you are treating the additives like they are categorical variables. This may indeed be the case but there is another possibility that you should think about. In my experience with paints, plastics, and various specialtiy compounds I’ve had the situation where we had a base formulation and the engineers wanted to try out a new list of additives. Sometimes the additives really were genuinely different in terms of chemical makeup and there was no way to rank order them in terms of composition, however, more often than not, the additive series represented one or two classes of compounds/compositions and the only thing that differentiated them was amounts of active ingredient. If you haven’t looked into the makeup of these additives you should. If it turns out that you have a situation where the different additives are just some kind of crude ladder with respect to levels of one or two active ingredients then you can reduce your run number and re-cast your coating sequence as a much simpler design. Also – are these additives mutually exclusive? If they are not then rather than trying a series of one-factor-at-a-time you would probably do better running a saturated design with a couple of your one factor experiments thrown in just to keep everyone happy.

    0
    #195299

    roarty
    Participant

    To begin with I should thank everyone for perservering with my problem and I shall try explain further my work as I now realise I did not provide sufficient background.

    I am developing a new epoxy paint which can be brush or spray applied to grit blasted steel in off-shore structures (oil rigs, wind turbines etc.). Yes the structure was pre-coated but over time the structure will require maintenance coating.

    Current epoxy paints are easily applied to wet steel and can even be applied by brush underwater. There is absolutely no problem with applying an even film of epoxy over steel which has been contaminated with sea salt. The problem with salt contamination arises once the epoxy film has cured and is in service. Although epoxy is a highly cross-linked polymer (like a fine mesh) even epoxy coatings will suffer from water ingress (hence corrosion/ loss of adhesion) this is because the water molecule is very small and will worm its way to the steel substrate over time (damn thermodynamics!).

    Once water reaches the steel, and if salt contamination is present, the salt will dissolve and the water ingress is accelerated due to the osmotic potential between the now highly saline solution beneath the epoxy film and the less concentrated saline solution outside. So it would be a good idea if I could add some ingredient in the epoxy coating which locks away the dissolved salt and so reduce the osmotic potential.

    When salt dissolves in water it forms positive and negative ions so I need to identify materials that I can add to my formulation which are effective at mopping up these pesky ions. My seven experimental formulations are identical to my control formulation but for the addition of materials which mop up negative ions or materials which mop up positive ions or a combination of these two material types.

    So if I take Robert’s advice; this is a two factor experimental design. The first catagoric factor has three levels (three chemically distinct negative ion mopping materials) and the second catagoric factor has two levels (two chemically distinct positive ion mopping materials). I’m still a little uneasy about “shunting the lurking variable over to the noise term”; the problem is that my noise term will be quite large as the difference in salt contamination between panels is large. I think I shall have to do several replications of my full factorial design which is more work than I originally intended but hey!

    So thanks for your suggestion Robert, I shall take your advice.

    Finally, Chris suggested doing MSA; This is an industry standard test and one which is recognised by our customers so even though I have never done a Gauge R&R, I am reluctant to devise my own non-standard test even if this test were to prove somewhat variable.

    Thanks everyone for your advice.

    0
    #195302

    Chris Seider
    Participant

    Just because you have a standard test doesn’t mean you have test variation that needs to be controlled.

    We (I’m sure others will pipe up even louder) suggest you confirm you don’t have much repeatability or reproducability or else you might come to the wrong conclusion on your tests. You might pick the wrong formulation because of test variability. Consider the example that formulation #2 gives a best result but it’s actually #3 that was measured with variation in the wrong direction.

    We have had many clients tell us the same thing as yourself and then was glad we insisted on conducting an MSA because operator effect. Even simple devices that everyone thinks works fine can be found to be imprecise. There are many reasons for this but miscalibration, misinstallation, poor training, poor maintenance, or lack of knowledge transfer are reasons why even a simple scale can fail an MSA.

    MSA’s are a basic tool in the six sigma methodology.

    0
Viewing 19 posts - 1 through 19 (of 19 total)

You must be logged in to reply to this topic.