DOE Planning - Make your Experiment a Sucess!

welcome to complexity made simple my name is paul allen and before we get on to today's video just a reminder of how you can support the channel please subscribe please leave comments on all the videos because that really does help but if you want to support the channel further there is a donation page please go to buy me a coffee and leave a donation which is always fantastically helpful keeps the channel keeps the channel moving and allows me to make all of this fresh material but if you want a textbook that contains all the good stuff from the videos please click on the link and buy drink tea and read the paper from lulu.com all of these things really help the channel thank you to all the people that continue to buy this textbook it really is fantastically helpful now let's get on to today's video okay welcome to complexity made simple my name is paul allen and the subject of today's little tutorial is how to properly how to properly plan a designed experiment um the phrase designed experiment is sometimes completely misunderstood people think that what this means is that somehow you had uh a thought before you went to the process of which experiments you were going to do and then you use your skill to decide which tests you actually plan into your experimental pattern uh so let me show you the sort of thing that that might mean so here's an experiment where someone's got five five variables that they'd like to understand and what they've done is they've decided beforehand so they've sat down and they've planned these eight tests out and they've decided beforehand and because they've planned them beforehand they will say this i designed experiments i used my skill to decide which tests it would be best to do now that is not a designed experiment a designed experiment which by the way is designed to give you knowledge it is designed to give you process knowledge follows a very particular pattern there is no choice in the pattern there is no skill in the pattern there is only the right mathematical way to approach the test so let me show you in a simple diagram what this would mean now we're going to think about three variables we're just going to think of time temperature and pressure you'll see why i only pick three variables in a minute but if we say time is going to be 5 to 10 seconds temperature is going to be 70 to 90 degrees and pressure is going to be 100 to 200 here's the pattern that gets tested we'll start with time it's going to be tested at the low and at the high only so we're going to test 5 and ten then what we're going to do is we're going to test temperature whilst those two tests are going on so this gives us four test combinations so this is now temperature here up to the side and obviously we go in 70 at the knot at the bottom 90 at the top 70 to 90 and then finally i'm going to put the third variable into the test and of course it creates this it creates this cube which is known as a design space and we've tested at the corners of the design space so we've tested at those eight places so this one here now would be pressure going backwards and pressure of course would have been tested at 100 at the low 200 at the high that is a designed experiment it is a pattern it is a mathematical pattern that gets the most amount of information from your tests and i mean i just put time temperature and pressure in there but you could put any variable in here uh it could be uh things like i don't know if it was a molding machine it could be hold time obviously it could be uh temperature could be temperature of the front tool could be temperature of the back tool um it could be injection time it could be injection pressure there's all sorts of variables that could go into this and it wouldn't matter what the process is this pattern if you're going to test three variables this is the pattern you're going to create a cube a design space and that is what a designed experiment is and it is nothing else if you've logically decided using your skill that's just a one factor test using your skill and by the way you'll find out very little using that type of testing only these types of tests give you true process knowledge so that's what a designed experiment is now how about going through and designing the experiment well in order to do this i have a template i don't just i don't just randomly make this up i don't do it in five or ten minutes i will often spend in fact i will definitely spend one day filling out this questionnaire one day it takes the design and experiment properly so let's look at the questions pretty straightforward what's your process what's your problem obviously what's wrong then it says okay what's your objective now i know what your real objective is to fix that but that's not that's not the way we want you to think you are doing a test you have the opportunity to find out incredible amounts of information just out of one test if you design it correctly so the the question is what do you want to learn what do you want to learn now this might be things like you might want to learn if there is a sweet spot with other outputs to the process so in other words you have a problem with one of the quality features of a machine you can't get it to land on target but of course nobody would thank you for curing one quality output and then messing up two others that's not an answer you need to find a sweet spot so there could be a sweet spot with three or four outputs that you're trying to learn um what else you could be trying to be cheaper so again you you might already have a relatively good process but you want to get the same results but maybe with some of the energy hungry variables turn down low and some of the less costly variables tuned in to substitute them that would be a great solution so there's more than just fix the problem yeah we want to learn stuff and you need to think that through clearly before you start then it says what's your start date and your end date what this basically means is how much time do you have to do this test time is always going to make you compromise it's going to make you compromise and so you need to know i've got the machine for the weekend i've got it for the day on a day shift how much time have i got and of course when you've designed the experiment you better make sure it fits in that space because if it doesn't you have a problem on your hands so how much time do you have then it says right what are the responses obviously the response with whatever's wrong in your problem but there should be more than one here key thing about the responses the outputs more than one so that what you can find is a sweet spot okay and then it says what type what type of data is the response now the response it must be variable data this is because of the sample size if you get past file data the sample size will turn into thousands so if you don't have a variable scale invent one okay so that's very important for the efficiency of the test then it says what's the anticipated range now a key question really here is can you measure the result can you measure the results the reason we're asking that question it goes back to the pattern that i showed you here what this pattern is basically going to do is it's going to put the three variables at certain times all at their low setting it's going to put the three variables all at their high setting so you're going to push the process around you're possibly going to put some settings in here that you've never used before and therefore what you might have you might have a jig here to measure the results in a very small window because that's all production is trying to um is trying to achieve but during the doe what we are trying to do is to push the process around i want the process to go on wild places so that i learn how that i made that happen and then i can learn how to hit targets so you need to just check if i'm going to go to extreme results do i have lots of lots of capability on my measuring equipment to just measure all the extreme results next point msa results key thing about an msa make sure you do one so we've got a good measurement system so we know what all the responses are we know that they're variable we know we can measure it and we know we can measure it really well because we've done an msa then he says okay what type of factors have you got now on this particular sheet i've only got six spaces simply because that's all that fitted on the sheet i think somewhere on here i've got i've got a page look that allows you to fill out many more but um just as a an indication of the sort of information that you're going to want what's the factor what type is it is it variable what is it attribute is it crn is it controlled or is it noise so if it's something like outside air temperature outside humidity that would be noise that's not controllable if you want to understand what noise is doing you're going to have to do a special robust designed experiment and the way the taguchi would have designed a robust experiment you'll have to look that up i'm not going to go into the detail of robust designs they are clever little tests they are running a very specific way though so if you want to learn to switch a variable off be robust to something and you're going to have to design the experiment to find that out then this is super important the range of interest so what's the range of interest in other words when we're designing the design space look my range of interest of these two figures the high and the low so for time it was five and ten um [Music] now this is where your skill comes to bear you have to decide the size of this box you have to decide where this box sits in space so i chose five and ten but i could have chose seven and twelve i could have chose ten and thirty five for time why did i choose five and ten that's the region that i considered the process would work and it's an it's a nice wide region so the key point about the range of interest is they should be as wide as possible but still make the results happen so if you turn the temperature down too low on a molding machine now plastic would come out i haven't got anything to measure that would be too far you need the window to be a nice big size but not in areas where the process would will not work and let me show you why this is we're going to wind on a variable at the low at the high we're going to get data it's going to be a little bit of variability in the data a little bit of noise in there okay so that was our that was our design space that was our choice there we chose those two points to test now in the case of the diagram i've just drawn the signal which is the rise in this thing is nice and big when you compare it to the noise because the noise is a little bit of variability that you get but imagine i've been a little bit too conservative and i've put the highs and lows much closer together now what would i say well i'd get the noise here and now there's the signal suddenly i can't see it the noise is overwhelming the signal that's why we want the highs and the lows the range of interest this is crucial the range of interest wants to be as wide as possible we want to turn the signals up for full blast and then once we see how the process works then we can hit targets then it says are we at two level what levels have we got two level or three level let's take a look at what that is now here look at two level i tested at the high i tested at the low and the relationship was a straight line but i could have tested it up the high tested it at the low and the relationship might have been curved well two points two levels is not good enough in order to pick the curve up i would do a test known as a three level test so do you think do you think the relationship is linear or do you think it's curved that's the levels then it says which variables do we think are going to interact and finally how are you going to measure these variables and this means how to set them how are you going to set them so for example let's go let's go back up to my uh let's go back up to my design space um you know let's say that on the the temperature on our machine actually we don't have degrees we we just have um we just have a dial yeah and we just know that if we turn it all that way it goes to the low and if we turn it around this way it goes to the heart we have a pointer on there but they've not given us a scale of any kind so the idea of going 70 and 90 on the temperature is almost impossible now what we have to be able to do because we're gonna we're gonna come at this with a pattern we have to be able to hit the values when we say we're hitting a high we have to hit it when we go away to the low we have to go to the same place we have to go to 70 we have to go to 90 and when we come back later and we want to hit 90 again we've got to come back to the same place now in this case what you might do is just put two dots on the scale and say you know set it to there set it to there that's perfectly fine for high and low it doesn't have to be complicated but you have to have decided how to do this these are not random tests the pattern that design space must be completed as designed so how are you going to set them so some variables will already have a scale you might have dials you might have led controls that you can easily set some of them might be your sharing cooling water and somebody says when we've got three machines on the cooling water gets pinched and the machine doesn't work as well when i've only got one machine on the cooling water works great how are you gonna set that you have to figure it out you can't just turn up on the day and just make this stuff up as you go along we have to design the standard operating procedure design the experiment know that we know all the variables now we're going to set them now we can go okay what's the sample size now the software the technique itself design of experiments is looking for 30 to 50 samples okay so you you need to assume they will all be defects you're going to push the process around to the edge of that design space there's a good chance that's not where the good product sits in those corners everything will be defective so you have to you have to take that on the chin um so sometimes this makes people panic with doing a designed experiment but typically you're doing the experiment because something's costing you money 30 30 defects at this point is normally cheap then you say which design types do you want to use so there's a design technique which is known as screening there's one which is known as modeling there's one i already mentioned which is robust design you could also have two level you can also go to three level okay so you need to decide what it is you need to know and how you think your process works and then you can decide which experiment out of that list you're going to pick then it says select the best approach now this is going to be a compromise cost is going to be on one side knowledge is going to be on the other you can make the test cheaper if you wish but if you do that you have to give up knowledge the cutting down sample size cutting down tests etc it saves you time it saves you money yes it does but it also you have to give up knowledge so you have to be prepared to give something up if you're going to if you're going to cut the corners if you really want to know how the machine works do the test as planned as published as recommended take 30 to 50 samples learn loads about your processes and make shared loads of money to me the best way to save money is to be at this end of this scale be it the knowledge end learn how your process works then it says can all the runs be randomized um [Music] randomization is a technique uh to protect you against variables that you didn't consider changing during the test so if i go back to my design space a second this these eight tests look like this i'll just put them in as the minus values represent a low the positive values represent a high so we'll just put the path and this is the pattern that describes this cube okay so there it is b c now you could just do them in is those eight tests they're in the order that they're published but if you do that you you do run a risk and you run a risk of a i'll call it rv which is a random variable a variable you haven't thought about let's say it's the temperature of the machine and the temperature of the machine starts to increase as you work down this pattern if the temperature of the machine has an effect its effect will actually show up as belonging to the first variable whichever the first variable is and we chose to put time in there so time gets a sort of bonus effect if you don't randomize okay so if you don't randomize if you randomize on the other hand you protect yourself from that effect now the question is of course randomization it costs money okay so typically your setups your setups are longer so you have to decide whether it's worth the longer setups whether you think you might have a problem with a random variable causing your results to be false personally my view is i don't randomize and i don't randomize because what i try to do is to remove all variability before i even attempt a designed experiment if you do that i typically find you don't have to randomize and that tends to be my procedure for getting around this but there's times in the past i've done it when i haven't been confident that i can remove all the variability and i've randomized the test you have to think you have to think about this do you think there might be random variables that could blow your test of course if so randomize then it says conduct the experiment record the results best advice i can give you be there the test the test has to happen as published it has to happen in this pattern here it has to happen like that if for some reason you get to a particular setting and you can't hit the pattern then you're going to have to decide on the fly what to do [Music] if new material arrives on the machine halfway through the test you need to make a note to that um maybe the machine breaks down halfway through the chest and a maintenance technician gets cool he has to replace something you need to make a no to that you need to just make a note of everything that's going on during the test because that they can all blow your results off course and the only way to do this is to be present don't give the test pattern to a technician and ask him to run those eight tests um he won't know what he's doing he won't know why the tests are the tests they are and if things change you won't think anything of it so be there and record everything that happens then do the analysis and a key point do a confirmation test do a confirmation test in other words what's a confirmation test you make a prediction this is when you really know how your process works you make a prediction and you hit it that is confirmation right there and that proves you've got knowledge out of this experiment that's what a doe does it gives you knowledge to be able to predict and therefore just hit targets um [Music] so in you you've proved nothing until you've done that confirmation test and this should be true by the way however you think you've solved the problem if you think your skill has solved the problem well do a confirmation test make a prediction of what's going to happen if you hit the prediction then you've done a pretty good job if you don't start again but the confirmation test is the true test of whether the doe has been successful or not if you confirm well done and then finally document your decision put it in your iso documents maybe make sure you keep company records of what you've done too many times too much testing is done informally and when that engineer or that scientist walks up the road he takes that skill and knowledge with him if you do a proper designed experiment and you document it properly and you keep the records that process you will always know how that process works and the engineer can walk and find a new job anytime he likes and you don't care because you've got company knowledge about how your processes work now there's the template that's that's what i do i always fill that out every time i do a designed experiment it takes me a day most of the day is spent on this page most of the day to be quite honest is spent on the table at the bottom here what are the factors you know you're going to get a team of people all your technicians and operators they're going to sit around talking about their pet theories about what makes this particular problem come and go there's going to be lots of discussion you will have to decide the range of interest that will challenge them enormously i've never been asked that question before um how to set and to come back to the highs and the lows i've never been asked to do that before so there's a there's a whole lot of challenging discussion to be had and often half of the day is spent on the table but spend one day designing an experiment properly design a proper design space like that get fantastic process knowledge and then smash your competition out of the water and make piles and piles of money [Music]

Channel: Paul Allen

Views: 163

Keywords: lean, six sigma, Six Sigma greenbelt training, Six Sigma Blackbelt Training, Shewhart, Juran, Deming, Taguchi, SPC, MSA, FMEA, DOE, X bar chart, Wheeler, Janam Sandhu, Mrnystrom, Gemba Academy, Full Factorial, Central Composite Design, Ronald Fisher, Hypothesis Test, p value, Histrogram, minitab, Pareto, multi-vari chart, https://youtu.be/QH984PnwRDE, https://youtu.be/f_fjqCpd67Q, https://youtu.be/AGJ1QYI2B4c, https://youtu.be/gsD8V2_eZ0A, https://youtu.be/mM6EyMvvAKk, quality hub india, simplilearn

Id: Z34q9sZJL9A

Length: 29min 6sec (1746 seconds)

Published: Mon Oct 25 2021