Moving from Dashboards to Decisions: Using Modeling to Improve Your Metrics (2026-US-30MP-2918)

Today, I'm going to talk about how you can do more with your data and move from dashboards to making real decisions. The way we're going to do that is by using modeling. A lot of times people will gather a lot of data, they create these beautiful dashboards, create amazing metrics, but they can go further, and they can get a lot more out of their data.

Quick introduction for myself. I'm Brad Foulkes. I run Gapz Analytics, it's a consulting firm where I have 20-plus years of engineering, process improvement, and data science experience. I've been using JMP for 14 years. A lot of what I do right now is trying to figure out how people or businesses that are collecting data can get more information and more insights out of what they're collecting today.

In today's manufacturing environment, there is so much data that's being collected. We know everything that's going on at the equipment level, when the machines are starting and stopping. We create these beautiful dashboards. We have so much information, but oftentimes, we don't really do a lot with it. It's a very backward-looking approach, and so we create these beautiful dashboards and then stop there.

When people are looking for what to work on and how to improve the metrics, they oftentimes rely on either gut feeling, tribal knowledge, or what's worked in the past and not really look at the data to see how are things improving, and where can I push towards; with the eventual goal being if you are already collecting all this data, how do you reach the point where you're doing smart manufacturing?

Smart manufacturing, you can't just start doing smart manufacturing. There's a roadmap, a path to get there. It starts with collecting data, creating visualizations, and building analytics. A lot of times, that's where people stop. They have their metrics, run charts, but they don't really do a lot with it. You can add in then machine learning and modeling, where you can reach the predictive analytics side where maybe you're doing predictive maintenance, prescriptive maintenance. You're doing a lot more with the data that you're collecting. That's when you finally reach that smart manufacturing side, where you're able to preemptively take action instead of react to what's going on.

Where we're going to focus on in this presentation though is around the machine learning side, assuming everyone's already created analytics, they already have metrics, they already have dashboards. But how do you go further? When I say metrics and dashboards, you can't just skip right to modeling, you do need some of these things. Some very common ones are overall equipment effectiveness, or OEE, with the components being availability, performance, and quality. You might have run charts, control charts, Paretos. You might have a lot of things. These are all incredibly useful, but they're backwards-looking. They tell you what happened in the past, they don't really focus on what's going on in the future.

I like to think of these as if you're taking your temperature when you're sick. When you take your temperature, you're not feeling great and you take your temperature, that's an indication that we know something is wrong. You have a high temperature, we know where we should focus maybe some of our efforts, but we don't really know exactly what to do. That's when you need to go to the doctor, you have more tests done. You do a lot more with the information that they're gathering. Same thing with using modeling and creating dashboards. Having an OEE metric that tells you something isn't right, it's fantastic, but it doesn't really tell you where to focus your efforts a lot of the time or what specifically you should be working on.

In terms of general process for building models, like I said, you can't just skip to a model. You need to collect the data, you're going to go through and clean your data where it might be you just need to understand what you have, look at null values, and make sure things are joining and matching correctly. That's typically where the dashboards come in. After that, that's when you can start building models. We'll talk about the different kinds of models that you can do.

But once you have models, then you can identify problem areas. You can identify, "What is it I should be focusing on?" Certainly, there are going to be some areas that you already know, "These are the things I should focus on now." But you also want to know, "What's going to be the area I should focus on in the future?" That's where you get to the forecasting solutions. You can do scenario modeling where you can say, "Well, what if I make an improvement?" Or, "What if I don't do anything, what will happen?" Try to have that forward-looking view of what's going on.

A lot of what we'll talk about today is how do you get to that forecasting side. Some of the benefits of this are you can identify worsening failure modes. Maybe it's not the failure mode that is the top of the Pareto or the worst thing right now, but it's something that's getting worse. Someone once referred to this to me as snorkeling. It's the failure mode that is just below the level of everything else where you're not really paying attention to it yet.

But if you have forecasting, you can do preventative maintenance. You can look forward and look towards the future to say, "What should I be doing? If I do this, what would the outcomes be?" It allows you to reduce unnecessary costs where you're not going in spending time on actions that aren't beneficial. Then I talked about the scenario modeling and asset management, where that allows you to plan for the future. This is not demand forecasting, where I believe a lot of people may think of it in those terms, but it does help with figuring that out.

As I said, this is a process, you can't just hop right in. The first thing would be cleaning manufacturing data. Now, a lot of manufacturing data… Let's hop to some manufacturing data here. Typically, what I would see from clients or working with people, what they would have is equipment data here that would show up in terms of events. What time did an event start? What time did the event end? What kind of event was it?

In this scenario here, we have start times, end times, and then what occurred. It ran for however many minutes. It ran from 7:00 AM to 8:35, and then it shut down. It was down for 30 minutes here. It would go through, and you would have that data captured. Then what sort of production, what sort of scrap. When you have this data, you need to make sure that it actually all makes sense. Does it all flow correctly? Do you have multiple events that are occurring? Are they accurate events? Do your production amounts and your scrap amounts look correct?

Some of that might be gathered from different data sources, so you might have to join different things together. When you're first calculating things— now, as you saw, I had just raw data—you're going to want to calculate durations and operating time. Operating time is how much run time has occurred for the equipment. Now, I'm sure many people have not captured their equipment from the very first moment that they turned it on, but you're going to want to just start to accumulate over time so that you can see how things change.

Other metrics you might be looking at: expected production amounts, so you might have an ideal amount that you're supposed to produce, or product-specific information, or categories. Categories might be day shift, how you're looking at things, or it could also be how you're grouping them. I worked with one client who was doing soda can bottling. When they first calculated values or calculated production, it was single can values, and then at the end, it was packages, which it might be groups of 6 or groups of 12. These are things that you need to make sure you're looking out for so that you can understand how the groupings are matching together.

Let's get into a brief example around cleaning some of your data. A lot of what you might end up with are two tables. This is what I would see frequently with customers. They would have one set of production time data where you'd have your start and stop time for the equipment, and it might say what happened, the downtime code. Here we've got Process 1, Process 5. These were reasons that the equipment shut down. Then they would have, in a whole separate data scheme in their database or in another system, the counts that were actually produced. This would be another table.

The first thing you're going to want to do when you're building models is how do you join some of this together? Because you want to be able to join your counts and your production time. Now, these are also the most basic data that you could be gathering. There may be other pieces of information: environmental data, you might include pressure or any sensors that are included in here. But for today, we're just going to talk about the most basic types of models that you could build. Then, of course, you can always go further.

Here, the first thing that we're going to do, we're going to update our production data because we want to join our counts data to there. We're going to join on a single column here. I'm just going to hit OK. What that is going to do is then join our counts into when they were produced. Now, as you may see, we have a starting quantity and a finish quantity, but it doesn't tell us when these were produced or the production rates at which these occurred.

When you have generalized data like this, you need to make some assumptions. You might assume the same production rate every single time. That is not always the case. If you have more granular data, it is more helpful and easier to see the production rates that occur. Same thing for scrap: you don't know when the scrap occurred, but you do have to make some assumptions around how this scrap actually occurs or when it occurs over time.

I'm going to jump ahead to the finished example of what you might actually do with all of this. In the end, when you join these tables together, you want to know what was being produced at what time. Like I was saying, you make some assumptions around scrap rate and starting quantity. What you can do is you can spread out the starting quantity over time and see what could have been produced at each individual time. This is just one way that you might want to prep your data and get it ready for how things are being produced. It does require making a few assumptions, but if your data is more granular, you are able to do things a little bit differently.

There we go. The next thing we're going to talk about is models. What models should you begin with? Like I said, we're starting out with a very simple example. These are time-to-event predictions, because what you're looking at is you have downtime events that are occurring. It could be that there was a clogged hose or some other downtime-causing event that is going to make your equipment stop, and you will want to know how frequently are these types of things happening.

The most simple thing that you're looking at is an event prediction model where you'll have a mean cumulative function. Right here in the graph, we have our mean cumulative functions where we're looking to see how frequently are these occurring over time. Now, if you have multiple pieces of equipment, this mean cumulative function calculation is a little bit different, but when you only have one piece of equipment. It's really just a count of time or a count of events over time; so however many operating minutes, however many events occurred.

You can use something called a nonhomogeneous Poisson process, all in JMP, to fit this type of data. The tool that we'll use in JMP to do this is a Recurrence Analysis. Now, if we wanted, we could do something a little more sophisticated, we could do time-series forecasting, we could do step-wise models. But we're keeping it simple for today.

Now, the second piece of information that you want to look at from, once again, our basic data is production and scrap rates. You want to try to understand what's happening when. Maybe there are some significant factors. Maybe you could create a model around the shift, the time of the day, the operator, some factor that could be driving production rate or scrap rate. You might have a team that is going great, and they're producing so much at a fantastic rate, well above normal, and you want to be able to take that into account. You don't always have that.

One other way that I like to use is something called a pick from list. What you would do is you would say, "The production rate or the scrap rate that occurred over time, I would select that as a possible rate that I could then use as a stand-in when I'm doing my modeling." We'll talk about that a little bit later when we have a more grander example. But right now, what we'll do is we'll get into an event prediction example and how you might actually build models out of this.

Let's say we started with our production data, and we had all of this information here: we've got our run time, our downtime. But really, if we're building a model, we don't actually care about the run time of it. Now, what we do care about is what was our operating time for all of these events. In this table here, we have all of those events. We have our unplanned downtime. Then I added in a few more columns. I've got the duration of the downtime, but then the operating time that this occurred.

The equipment ran for 95 minutes and then had a 30-minute shutdown. It ran for another period of time, and it's up to 132 minutes. As you can see, this is always going up. Now, the other important thing to notice is that I have all sorts of events here, the operating time is always still going up, because this is one system. Even though there are different causes or different failure causes, it's still one system, so the counts are always still going up.

Let's build a model here. We're going to go to Reliability and Survival, and Recurrence Analysis. For our model, the operating time is our age, because that's the age of the equipment. For this example, we're going to put in event types as our label. Now, if you had multiple systems, you'd put multiple systems in here, but we're looking at one piece of equipment. The other thing we're going to do is our failure mode. We'll add that as our cause or failure mode.

Now, you don't always need to do this, but one thing I do like to do is what was the end time stamp or end age? For here, I'm going to put in 65030, and then we're going to create our model. The recurrence analysis that you get here… You can see this might look familiar, this is our MCF plot, our mean cumulative function plot. It's broken out all of the different failure modes here that we have in our dataset. We want to build a model, we can come up to the red triangle, Fit Model.

Now, we can select from a few different ones, but we're going to stick with the default one for now, the nonhomogeneous Poisson process. It's a very flexible model, a very simple model too, that is very easy to understand. It has just a couple factors in there. If you had other features in here, you could add those in to say, "How do I want to improve my model by adding it in as a scale or a shape effect?" We're not going to worry about that today. We're just going to run our model. What you can see is that it produced a line of best fit for every one of these MCF lines.

We have all the parameters down here. We can do different things: we could look at a profiler, we can save our formula. We can do all sorts of things here with this. But this is where we would start to build our models. Once we have these models, then we can do that forecasting. We can do that future look that I was talking about.

We did our time-to-event. The last thing or the last example I'm going to show you is what happens if you tie this all together? I have a sample script that I've run where we have example data. Let's use our example data. This would be where you have an add-in solution in JMP that you think to yourself, "These are analyses that I want to do frequently, so let me create something that allows me to create a report and do some of this modeling." I'm going to throw that in there for our production rates and our scrap rates, set what events are my planned and unplanned, downtime. We'll leave that as the ideal rate.

With this script, it goes through and produces a series of charts and does a lot of the modeling that you're interested in. Taking a second working in the background here. The first thing it produces is a list of your descriptive statistics. Like I said, you don't want to just skip over your metrics and your analytics. You want to make sure you see these and look at these because there could be some very valuable information in here. You have descriptive statistics, just run charts for the various items, and you can see where you are at.

Your OEE, in this example, is just slightly above 70%, availability at 80, and so on. You also want to know your Pareto of events. When I look here, the first thing I see is production P1, P2, and mechanical M5 and M6 are my top events. These are the most frequently occurring events that have a significant amount of downtime. P1, 139 events in the dataset for roughly about a quarter of an hour or so. Mechanical M5, 86 for about a third of an hour. Maybe I want to look at the duration of events that are occurring for every one of these. I might come in here and say, "Let me look at some of these and just see what's going on with my data."

Nothing all that crazy here. This gives you a chance to look at your data and say, "What's going on? Is this normally distributed? Is there a bimodal dataset?" If there is some bimodal aspect to your data, you're going to maybe want to separate those out. Maybe you have two different types of events that are occurring, but they're all being marked under one bucket. This is typically what most people have and where most people stop.

Now, if you go into the modeling side, you can come in here and say, "I've got my mean cumulative functions. What if I go then and create all of my models? If I look at these, I want to know which ones of these are getting better and which ones of these are getting worse." Now, this doesn't focus on the downtime duration so much as who's getting better or worse. Mechanical M2, based on the model that we have, it looks like it's getting worse over time. If I look at M2, something I see here, I've got 19 events, so it's not one of the top event values here, but you can see this exponential shape to it. As the machine is operating, it is getting worse, and it's going to continue getting worse if I don't do anything about this.

One that's getting better here is M3. Let's take a look at M3. We can see this is leveling off. Same thing with the counts, there's only 48 of them. But it looks like there were a lot in the early days, and now there are big gaps. It's starting to get better over time.

Something else you want to look at is the actual versus predicted values. This is just a calculation that you can do with your model. Now, the one that I would be concerned about here, some of these I'm going to overpredict, some of these I'm going to underpredict. M4, I'm underpredicting the number of events that would occur by six. It's pretty large. If I come over here, and I look at M4, what I see is I've got a fairly consistent line, and then towards the end of the data that I was collecting, I had this huge spike in events.

This would be something that would completely fly under the radar if you're just looking at Pareto charts. You would never really focus on M4 here because you've got these five other bars that you would tackle first. But looking at this, you can come in and say, "What happened there? Is this about to explode? Is this changing in behavior? Do I need to do something? Why are we having so many events that we want to dig into?" This does not negate P1 and P2 being the highest bars here, but it gives you another area that you want to look at for what is the next thing that might be occurring.

Then you can break this all down and say, "If I were to look at this, where should I really focus some of my time?" Once again, this is looking at your next thing? What would be the biggest thing to tackle as your future events? I like to break these into four categories where you've got your changing mean time between failure, your average downtime. Anything in Category 1 has a high downtime, and the mean time between failure is getting worse. Over here, the clear winner is our Mechanical M2, where the mean time between failure is getting a lot worse over time. Conversely, Mechanical M3, getting better over time, probably don't need to focus on this.

You can do something like this where you're looking at, "Where should I focus my efforts here?" Now, all of this leads to the scenario modeling side. Now, we're going to run a quick simulation while I talk here. The simulation is taking all of the models that we've built now, and it's running in the background to say, "What will happen in something?" Here I've said in the next 30 days. You could set yours to 90 days, 6 months, whatever you wanted to do. But you'd be able to take a look at this and see what actually happened over time.

Now, this should pop into the window here. I'm not sure why it didn't. But what you can then do is say, "Maybe I want to know what would happen if something got better." If we've got our window here, and we're going to improve this by, let's say, 20%, we take this. Then this allows us to see how do things improve over time. For this one failure mode, what impact would we expect to see if we made this improvement?

Now, your improvements might be that you're adding more people to the line, you're changing a piece of equipment. We made an improvement here on Mechanical M2. What we were able to do is say, "How much of an impact did this make to our various factors?" We only ran it for 30 days, but we can see this is what happened to availability over time. If we made this sort of change, we would see an improvement in availability, on average, by about 0.7%. Now, depending on how much you're operating, that's going to translate to a certain number of hours or minutes in a day. But it allows you to quantify what impact these different things would make.

Additionally, you can look at OEE and the improvement that this could make. Now, we're only changing one factor, so depending on what it is that you're doing, you could have different factors, you might be able to change the production rate. This is one solution for a very simple set of data.

In summary here, models are the next step that a lot of manufacturing can take to get to smart manufacturing. One step that you can take to get to smart manufacturing: start simple with models and then you can always make them more complex. As you have more factors, you can add them in, you can figure out what's significant, what are drivers, and you always can make them more complex a lot easier than making them less complex.

Always gather your insights before taking actions. If we went and just made changes and then said, "We made these changes. What impact should we expect?" You don't want to do that. You want to look forward, look towards the future and say, "What improvement should I expect? Is this even going to pay off? Do I think it will pay off?" Forecasting is that tool that allows you to identify what impact, where to focus your efforts, what's going to be worthwhile.

Thank you very much, everyone. That's all I have for you.