How to Use Statistics as a Common Language – and How to Improve Process Quality by Doing So (2021-EU-30MP-788)

[#1]

Level: Intermediate

Marcus Soerensen, Head of Quality & Six Sigma, Envases

This case study will show how the quality of a process was improved by using statistics as a common language between various departments in the company. What started out as a "confused" project with a very limited set of data turned out to be a successful project as the team began structured and data-oriented progress using JMP. This case illustrates how various departments can work together using data and statistics as the foundation for process improvements.

Auto-generated transcript...

Speaker	Transcript
Marcus	Welcome
	to this presentation here, using SAS JMP.
	The topic for this presentation is how to use statistics as a common language and how to improve process quality by doing so.
	I'm working as a quality manager for a company called Envases. Envases is a huge company, a
	global company making cans for the food industry. You can see some of the products here in the in the Left corner.
	Some of the products may be familiar to you. It's cans you can buy in the supermarkets, we have meat, have fish, we have milk powder, we have also juice and so on in the cans.
	We have manufacturing facilities in
	Mexico and in Europe and the
	topic that I want to discuss here today is related to our production facilities in Denmark.
	Basically what I'm going to present here is a problem that has been solved using SAS JMP as one of the tools.
	And I'm going to spend some time in the JMP to to show some of the functionalities and to give an idea of how SAS JMP can can be used to work with these kind of problems, which I'm quite sure is very familiar to many of the people, seeing this presentation.
	Well, going back to last year,
	we had some issues in our production lines and the people in the production came to me. I'm the head of quality and I'm dealing with these kind of problems that came to my desk saying, Marcus, we have some some tin dust in our production lines.
	Tin is the material that we're using for cans, so tin dust as such it's not
	it's not around things to see in our production, but people were complaining that they could see too much tin dust in our production lines.
	And, as a result of that, they need to clean the lines all the time.
	That was, like the beginning of the problem. We didn't have any data to to show this. We didn't have any paper to go through. We didn't have any Excel sheets. We didn't have any JMP files to to look into. We just had some opinions from the people saying we have too much
	tin dust on our production line, and we need to clean all the time.
	And to make it even up a bit
	more confusing, but people were saying we have seen this many times before, and some people were saying that we have never seen this before.
	I started to speak with the people on the on the line, saying, is this new to you? Has this happened this week or was it last week? Or what about last year? And I got a lot of different opinions and a lot of
	inputs, and also people were not telling me the same thing, so I was a bit confused about how could we get started with this project.
	But anyway, we needed to get started.
	So the first thing I did
	was to set the team, saying,
	people here in the organization will need to set a team. Who should be involved in team? And we pinpointed some some people in a in the production, some technical people, some people from the operation, some people for the maintenance department and so on, relevant people for the project.
	And the structure that we have used here to solve the problem is inspired from the Six Sigma DMAIC, maybe some of you know it already.
	DMAIC is about defining the problem, measuring the problem, analyzing the reason for the problem, improve and then control, if you actually succeeded with the with the solution that you came up with.
	One of the idea with this, DMAIC, and one of the idea, also for me, using the structure here is to define and measure the problem, because sometimes you
	may be thinking you have a problem, but by starting to measure the problem you may realize that what you think was a problem is actually not a problem at all.
	So there's no need to to continue solving the problem that is actually not existing, and that was what I told people saying, Okay, first of all, let's measure the problem. Do we even have a problem, or is it just a stomach feeling that we are having in the production?
	That was the first step.
	So we went to the to the production line ???, going to the line to see,
	can we have an idea of the process, maybe we can even see where the problem occurs.
	The process is is is rather simple. I've tried to simplify it here in this very simple drawing,
	just to give you an idea. We have tin sheets coming in to a conveyor and then we have a piston
	making the lid for the cans, so the process is rather simple.
	As a result of this piston and of making these lids, we can see that we can collect some dust on the conveyor, and this is where the operator needs to clean once in a while.
	And we could see by doing the ??? that after 25,000 lids, we could collect about or more than two kilograms of tin dust.
	And that was when we needed to clean the line. So that was like our baseline, say when we had produced 25,000 lids equal to more than two kilograms, we need to clean the line, and that was
	not acceptable because cleaning is taking up production time and having reduced production time we cannot produce the lids that we want to produce. That means that we maybe will be late for the delivery, or maybe we are not able to deliver at all.
	So we had, all of us, an interest of reducing this cleaning and we could see that we can reduce your cleaning if we can reduce the tin dust produced at the line.
	So the problem was pretty simple when you when we have collected the
	the
	data in kilograms here.
	Remember that you can see the back here on the picture, this is the dust that we could collect using a vacuum cleaner after 25,000 lids. In the beginning, when people came to me we didn't even have a baggage with the dust, it was just
	it was just by watching, we could express the problem that we were having. Now at least we could see that this is the problem that we're having. We have too much dust and this is too much dust, because we need to clean it. So we have to reduce the
	kilograms after 25,000 lids produced, that was the success criteria of the project.
	By collecting the right people, we took a like a workshop, put all the people in a room saying we have a problem here. After 25,000 lids we produce more than two kilograms of dust. What is the reason for that?
	And then we did a brainstorming, saying look at the process, looking at what is coming in the process, what is coming out the process. Can any of these variables explain the recent for the tin dust that we see?
	So step one was to define the variables. And you can see, I have marked here in yellow what the team expected to be some of the root causes for the tin dust that we saw.
	We started with the input that since it's coming in, we know that the tin sheet, the thickness of the tin sheet can can
	can vary, so we can have some thicker tin sheet coming in and some
	not so thin thin sheet coming in.
	And we also, we could see that the coating of the piston could be a reason for the tin coat.
	We could see, compared to other lines that had some coat, didn't have the problem in the same scale.
	We could also see that the measurements of the piston, we have four different measurements on the piston that could explain the reason for the tin dust.
	We have never tried this out, so this was just on the paper, so this could be the reason but let's try to find out.
	Last year or two years ago, when we have a problem like this,
	the approach would normally be that we were trying different things out, so we could try and make the thickness could see if this could change anything, the coat. But here we would like to combine all the variables in one experiment simply to to speed up the process.
	So we set up a design of experiments, a DOE.
	Over time we have to change this a bit, so it could reflect the reality that we have and also the allocated production times that we could use for the experiment.
	Setting up the experiments and defining the variables was not a difficult task. It took maybe a couple of hours. Setting up the DOE was not difficult, we did that in SAS JMP. But executing the DOE was the tricky part because it took a long time; it took about a week.
	So we need to plan to take out the machine and then we did the trial for about five days. And simply what we did, we produced 25,000 lids using one kind of setting and then 25,000 lids using another kind of setting and so on.
	And then, after the week, we analyzed the results, and then we concluded based on these results.
	Let me try to show you
	what we did in SAS JMP.
	You can see here this, just by looking at the numbers, we could see that this is a huge progress since our starting point. We started by just having people say we have too much dust and we need to clean all the time. Now at least we have some number a number...numbers on the on the tin.
	You can see here, we have the tin dust. This is the
	produced tin dust after 25,000 lids, and we also have different settings of the thickness of the material coming into the line.
	And we have the four different measurements of the piston here. And we have a statement, has it been coated or has not been coated. So we have different kind of pistons that we were trying out.
	This is rather easy for people to understand. They could they could see how much tin dust do we have if the thickness of the material coming in, is 5.74, if the measurement is 1.47 and so on.
	So this is a huge step from coming from just watching to actually have some real numbers behind the working set that we were having in the beginning.
	So just collecting the number here was a huge progress from our starting point, but the idea was to use the number to see can we explain
	the reason for the tin dust based on the thickness of the material coming in, the four different measurements of the piston, and if the piston has been coated or not.
	We're using some of the tools in SAS JMP and one of the tools that we're using a lot here is the fit model. The fit model explains if there will be any relationships between your responses and your variables.
	And up here we have the response.
	This is our tin dust.
	Here we have the model, so we would like to see if the thickness of the material effect
	the 10 dust, the measurement of the piston,
	and if the coating of piston would have any impact.
	Running the model here
	and saying, try to combine the different variables and tell me what will have the highest impact on our tin dust.
	And this is basically the results that we got out of it.
	We could see here we have what we're using the p value for us to guide us if this makes sense for us. The coat is low and the p value meaning, well, it seems like we have a significant
	relationship between the coat and our tin dust. We also believe that the thickness of the material will have a
	relationship with our tin dust, and we believe that the Measurement 3 will have some kind of relationship with the tin dust. This mean also that the Measurement 1, 2 and 4 don't seem to be significant when we talk about the tin dust.
	Remember that we were starting from just working without any kind of number, so now we were talking about P values and how this can help us.
	And this was actually quite easy for us to to interpret it, and people did understand, okay significant means that this maybe is not a coincidence.
	And by using the right people, we could verify this makes sense for them as well, so it seems like the coating can have an impact on the tin dust.
	And the technic... technical staff were saying, yeah, it makes sense that the coating will impact on the tin dust, because we have seen this on other lines.
	And the thickness could be verified makes sense and the Measurement 3 could make sense, so we started to believe that this is
	some good guidelines for us, but we need to see yeah, we can see that the coat seems to be significant, but is it with or without coating that is relevant for us?
	So we expanded this...
	the fit model here and
	you can see here in the in the profiler how the different variables will impact the tin dust. You can see the tin dust here to the left.
	And you can see the thickness of the sheet coming in the process here, the Measurement 3 and then the coat. And then we could try to simulate if we have to coating, if we have Measurement 3 on.
	what would be the expected kilograms coming out of the process?
	Here it's saying we can expect 0.0057. We also have a confidence interval here, but we can expect that this will be the, the number of tin dust coming out of the process.
	Then look and see what if we have a piston that is not coated. Can see that will change significance, and it will be higher.
	And we know from our started that around 0.2 will be like the game changer if we have more than 0.2 kilograms in tin dust, we need to clean, so we want to be lower than 0.2.
	And what we did to go even further here, because we know that the thickness of the material was very difficult for us to control, this is specified and there will be some variation within the thickness, which would be very difficult for us to change.
	So we needed to have a very robust process, saying we need to keep the thickness flexible.
	But what we can control is the Measurement 3 and the coating.
	So we expanded this profiler to
	the simulator so that we could simulate what if the thickness will change
	with some specified standard deviation.
	So we're saying we know the thickness can can change.
	We know that we have a mean around 5.595 and we have a standard deviation of this material equal to 0.058. We could change this later on. We want to fix the measurement and we want to fix the coating, having
	the coating on the piston.
	We also know that our target is not higher than 0.2 so we could add a target in here.
	And then we could simulate.
	The Measurement 3 will be fixed, the coat will be fixed, but the thickness will change over time.
	Then we can simulate if we want 5,000 sheets in the process, what could we expect to see
	in the tin dust? And then we could simulate. You can see if we have
	a tin sheet coming in having a thickness of about 5.6, the Measurement 3 at 1 and with a coat, we could expect a very good result. And you can see, we have the red line here at 2...
	sorry, 0.2, and we can also see hfere that the rate of defect, meaning that rate of measurements higher than 0.1 would be 0, so this is good for us.
	You can also see that if we then change it, the Measurement 3, not at 1,
	but to 2.5.
	And with the simulation again.
	We will be at
	a slightly higher
	tin dust amount. If we, on top of that,
	sorry,
	change it to a piston with no coating and run 5,000 sheet plate, we could expect a very poor response.
	And we could see that
	the setup that we had in our production line before we started these changes were pistons without any coat. So this was very new to us, and it was very exciting for us to see that we can actually see what could control the the tin dust and we can even control it ourselves.
	So, based on the simulations that we have here, we decided to have
	a
	Measurement 3 on 2 and then a piston with a coating,
	meaning that we will have a very robust process that gets...that can handle the variation that we have in our thickness of the sheets.
	Then we could also try to to to simulate thing, what if this deviation will not be 0.058 but will be
	1.5.
	You see, this will be too much, so this is some of the agreements that we have with our suppliers that they need to keep the standard deviation at a specific level because then we will have
	a robust process.
	So this was like a an eye opener for all of us since this is very, very good picture for us, and it hasn't cost us a lot. The only thing it has cost us was the experiment that we set for five days.
	So this was good news for us.
	And so here on number five, we concluded saying, Okay, now we know what should be improved, we need to coat the tools or the piston and we need to adjust the measurement to 2.0.
	It was coming from about 3.0. So this was ??? an improvement, and this is a very simple picture from the profiler,
	showing the relationship between our tin coat and the different variables that we expected. And we did a control saying let's try to change
	our processes and then produce 25,000 more lids and then you can see the comparison. This is was our like our baseline. You can see, the small dust here in the baggage before we change the process, and you can see the 25,000
	of the tin dust after producing 25,000 lids after the change and we didn't see any tin dust here.
	So this was like...it was very good proof of concept for us. It was one of the first projects that we did using SAS JMP and it was a very good proof of concept and people did really rely on this way of doing problem solving.
	What we learn here is the use of data, it shows benefit for us, because then we have something in common, something that we can relate to everybody,
	instead of having different opinions that is very difficult for us to quantify, so the use of this was very helpful for all of us.
	And using the right people, the technical people, people
	with a knowledge from from the line, also people who can use like a statistical software as JMP understand how to set up
	an experiment, understand how to do a fit model, regression models and so on. And then the use of SAS JMP is truly powerful. We could not have done this, like in Excel because they don't have the tools.
	And then, using a structured process like they make us very powerful for us, so this was a very good learning for us, and this is something that we have implemented in many projects afterwards with very good response.

[Lu]

Nice presentation Marcus. Good work floor example of Plan Do Ckeck Act (PDCA) with the help of JMPpro dataanalytics.

One question. I want to set up a custom DOE myself to optimize the heating process of the microwave to warm baby feeding. I have a model with 4 categorical factors and 3 continuous of which 2 continuous) factors (warming time & volume of milk) are easy to change. I saw in the DOE table of your presentation that the continuous factors were split up into several random levels at each categorical factor. So far I was not able to get this done. My DOE table only use the under and upper limit value of the continuous factors instead of random levels.

How did you succeed?

Hope you can be of help .

With regards,

Lu,

Belgium