Carl Zeiss Vision Care produces custom-ordered lenses. They get raw material, form the glass, run it through finishing operations and quality controls.

Their analytics and rules engine monitors everything: The labor, the production, the quality, the equipment, and the environment.

The data indicates when the factory’s about to fail, giving them a window to schedule maintenance, reschedule, and look for operational excellence.

“This is a current example of smart manufacturing in medical devices,” 10x for ENGINEERS presenter Srihari Yamanoor says. “In some places, it’s like, ‘Oh, this doesn’t happen in our industry.’ But it is happening. Right now.”

Watch his presentation and learn about the inevitable.

Srihari Yamanoor: Good morning, everyone. So, my name is Srihari, but you can call me Sri.

My background – I’m a mechanical engineer as well. I’ve been working medical devices for about 12 years now, but I also have a background in CAD/CAM and in Quality Assurance. With my brother, I do some digital health products now, and we’re actually building AI into our products.

So, you know, there are two different types of hats I wear. At work, I do women’s health, and I came off a big project where we took a redesign and transferred that to manufacturing, and so we’re seeing some of the pains we have when we don’t use automation. So, I’m hoping that this is something that is of interest to people.

I have 25 minutes. It’s so little time to cover so many things. I’ve tried to blog about some of this stuff. I’m also going to upload these slides. I try to keep the wording on my slides down, but there’ll be some footnotes in it after I upload it. I’m trying to get a jump start here, so we can have a discussion afterwards.

So, we know this – the things that a factory looks for never change. Hopefully, the first thing we look for is safety. Then, we’re looking for the quality of our products, productivity comes afterwards, and, of course, the focus of all of that is profitability.

I’m sure we always want to stay ahead of competition. Whether you automate your factory, whether it’s all hands-on operation, whatever way you function, these are the primary codes.

Advantages for medical device manufacturers

But from the medical devices industry standpoint, what are we looking for?

Obviously, regulatory compliance is at the top of what we are doing. Right? Quality is the next thing. But in terms of a smart factory or a smart factory floor, whatever you want to call it, there are other advantages we get.

Some of the stuff we struggle with — instead of manual data entry, trying to collect automatic logs, automating the production, if you will. This is an example I will give as well – customizing the products for the patient population or the doctors that use them is another big goal and then, of course, the integration so that you can handle servicing and CAPA better.

So, these are the universal goals that drive us towards automation or towards a better factory flow. I just wanted to lay that out.

How we got here – the four revolutions

That was a brief overlook at how we got here – first came mechanization, then mass production with Ford and everything, and, of course, we did do automation over the last few decades. So, what we are calling smart factory today is different, and we call them cyber-physical systems, IoTs, so many other terms.

I wanted to see if we can discuss what that means for us. So, how did we get here? The way we progressed here is we went from computation.

There was a recent study that I came across. They claimed that 80 percent of all the recent advances in AI came through the improvements in computation over the last couple of decades. Computation is not just the hardware, but also the sophistication of the algorithms.

In fact, if you’ve been keeping in touch with what’s been going on with the Siri and Android Voice, a lot of it came through deep learning and improvements in neural networks that came through because of scientists like Hinton, who now works for Google, who did make these advances over the last couple of decades.

And, of course, we now have the ability to put sensors on everything, so that’s been another advantage, and that gives us immense amounts of data. And then, the question becomes, “How do we analyze this data?” And that’s where AI comes in.

When we say AI, we’re talking about general machine learning. Right now, we don’t have the artificial general intelligence. So, what does it all mean, in terms of the objectives and what we’re trying to do?

Smart manufacturing

Now, this is not the full set of definitions of what would be a smart factory, but it would have many of these elements – artificial intelligence or machine learning; robotics and automation; human collaboration with machines, that’s what I mean there; customization, like we talked about; and, in some cases, novel manufacturing like 3D printing; and then, a few years from now, hopefully, 4D printing; self-assembly, self-correction, that sort of thing.

We have that in software now. We will be expecting to push that into hardware. “Oh. I am defective, I’m going to fix myself before I get off the floor.” Right?

And, then, integrated quality control, of course. But if you think that’s exciting, that’s just today.

Tomorrow, we will have artificial general intelligence, something that’s able to set its own goals, insofar saying, “I want operational efficiency. I want scheduling efficiency. I want improvements in this product and that product.” It’s able to find out what it needs to fix – and fix it!

We go to nearly fully automated with minimal involvement from us. The sensors will keep getting better.

Today, a sensor is hardwired. It only measures pressure. It only measures voltage.

Tomorrow, sensors will be different, flexible electronics. They’ll have their own intelligence and the ability to change what they measure, and, of course, we are expecting other improvements in manufacturing tomorrow.

Sample data sources

That said, what do we mean by data? Like what sources of data can we have you have? We have data coming from:

Incoming inspection. We already have that. We already do that.

In-process inspection.

Testing. Yesterday, we had a speaker talking about testing. Testing is an integral part of medical device design. We can take that data.

Equipment monitoring – so if you use cleanroom operations, we do that. So, you’re always monitoring your key room. You’re always monitoring the temperature, the pressure, humidity, particle levels, and all of that stuff.

Production data.

Field data and planning.

One of the things I didn’t include – this is a new concept that’s taking on – is the digital twin.

That’s something that is going to be the future. You’ll have your entire factory floor simulated, and you can run simulations on what could go wrong, how things can be improved. All of that data feeds back.

So, these are all sources of data. We have immense amounts of data coming in that are the capabilities, so we need something to analyze all that data and that’s where AI will come in.

Before that, how would we collect this data, and how would we send it forward for analysis? That’s where cloud computing or edge computing comes in. They also call it node computing. Cisco wants to call it fog computing, so they can have their own trademark.

Connectivity, storage, computation

What does it mean?

We have computation on one side. We have the network connectivity to upload the data somewhere else, and the ability to store it so that you can do both short-term analysis and long-term analysis.

So, then, the question might come, “Why don’t I just upload everything to the cloud? And why don’t I just go ahead and analyze everything all at the same time?”

Well, you’re going to have immense amounts of data. You already have immense amounts of data.

Whether you’re doing the analysis now or whether you plan to do it five years from now or 10 years from now, the data is going to be immense in volume and quantity. That will choke your network.

It’s one thing to expect long-term analyses, looking at long-term production, run schedules, and so on. But when you need the information right now, by the time it uploads to the cloud, by the time the analysis is done elsewhere, by the time it comes back to you, it is too late.

That is where edge computing comes in, where you will use that network connectivity, that storage, that internet within your organization and get a rapid response. You will have a robust set of operations. Plus, think about this – this doesn’t exactly apply to medical devices, but, for military applications and things like that, they don’t necessarily want their data ever leaving the floor. Right?

So, for all those purposes, edge computing has – or fog computing, in this case, whatever the name is – that’s where the advantage is coming.

Smart manufacturing example – Carl Zeiss Vision Care

I was looking for very good example. There’s a lot of hype and articles, so this is an excellent, in-process, smart-flow operations example I pulled off.

That’s these guys, the DXC people who created the analytics engine for Carl Zeiss. Carl Zeiss Vision Care produces these lenses, and they customize them to order. So they have a series of processes.

I know it’s a little hard. This is my busiest slide, and I apologize for that. But when you download them or you go to that link, which is in the notes, you can see they’ve described the process.

They get the raw material, they form the glass, and then it goes through a series of finishing operations and a series of quality controls. Now what they did for the factory is these guys set up the analytics and the rules engine, because this went through and started monitoring everything.

They track the labor. They track the production. They track the quality. They track their equipment, the environment, and all of the data allows them to know when the factory’s about to fail, allows them to know when to schedule maintenance, allows them to know how to be able to customize and reschedule and look for operational excellence.

So, this is a current example of smart manufacturing being used in medical devices. In some of the places I’ve worked at, there’s a lot of inertia, and it’s like, “Oh, this doesn’t happen in our industry and so on.”

But this is happening. Tomorrow is already here for us. That’s what that example shows us.

Industrial Internet of Things

I wanted to touch base a little bit more on what the possibilities are, so I found this nice example.

There are many things we can measure, right? So we measure them, but one of the most important outcomes of that should be actionable information. That’s what we’re looking for, and that’s what this represents.

So, I’m monitoring my equipment. I’m monitoring my labor. I’m monitoring my supply chain, my inventory, and all of that stuff.

What do I do with it? You want to be able to change your maintenance plans. For example, one of the things we struggle with is how long should something cure. After I open a set of chemicals, one of the things we struggle with is we have all these specially curing adhesives. We just use this blind sort of “we’ll open it for 15 days and then throw it away.”

You know how much wastage is generated by things like that? Or when you stop your machine once every 30 days to do maintenance on it, that’s wastage. You may not need to stop every 30 days, and, maybe, 30 days is too long.

So we want to be able to get information up-to-date and accurate, and that’s the purpose of the IoT. And it is not hype. It is already here. We have an excellent booth over here as well.

Greenfield vs. Brownfield

Another commonly discussed problem is “should I go start a new factory or should I go ahead and do an upgrade?” Now, sometimes you don’t choices, and, sometimes you do.

If you’re a startup, you’re in a good place to start operations from the ground. You will be using contractors with components and some of the stuff that comes from elsewhere, but your assembly and your operations will still be in-house.

Or if you’re introducing new product lines or if you’re creating a second factory floor or a third factory floor, that’s a great place to start greenfield.

Brownfield is when you have an existing facility which has its own regulatory approvals, licensing from your regulatory agencies, and you can’t really go start a new thing, then you do upgrades. And then the question is, “What are the challenges?”

There are the financial challenges, and they’re slightly different but comparable. If you’re doing a start-up, it’s your start-up cost. If you’re doing an upgrade, it’s your upgrade cost. Whereas what are the uncertainties?

The uncertainty for a greenfield is, “Will this work? Is this the correct way to do something?”

Whereas, for a brownfield, your uncertainty is, “Can this be upgraded? Will this be compatible with the newer technologies or is our network so old, it can’t work with the cloud service that we want to implement on our floor?”

What I’m trying to drive at is, no matter how you look at it, there are no excuses and the difficulties you’re going to have are comparable.

Handling data with care

So, before I switch to AI for a little bit, I wanted to just say one thing: It is very important to know what data you’re collecting. All right?

We can collect a lot of data, but it is important to what we’re doing with it, how we are analyzing it, and also important to know how it can be misused.

There are two of many ways things can be misused. One primary way is you can end up with micromanagers who use this data to do all kinds of stuff, so you got to be careful how to prevent that from happening.

I wrote a couple of posts about that. But the other thing to remember, and this is key, is security. Industrial espionage is just as old as industry is, which means that now that you’re collecting all this data, you don’t want to put it in a nice package for someone to steal from you and know what you’re doing and how you’re doing it.

So those are some important considerations to think about when we’re moving towards a smart plan.

What is artificial intelligence?

So, this is one of my favorite slides about AI.

I don’t know if you guys have come across this story before, but eleven blind men walk up to an elephant. They all touched a different part of the elephant, and they come up with a different message. So, what I wanted to do was draw a quick baseline on what it means when we say AI in the factory or AI on the floor.

If you want just the 32,000-foot overview just for being able to say what it is, it’s a discipline where it’s a learning system. You always give it data. It learns from the data and then you apply it. And it happens through computer science where your algorithms come from, and engineering, which is what we do are our floor.

This is another slide that just lays out the different terminologies. AI stands at the broadest level. This is from a famous reference text.

We don’t have “general AI.”

Whatever we define as AI is “narrow intelligence,” which is basically composed of machine learning, and I’ll describe that in a little bit more detail; or “representational learning” where it learns the features by itself; or deep learning, which is really how we are able to get voices on our phone, machine vision, and all of that stuff now.

So, that’s sort of broadly defining the field of AI. And when most people say AI now, and most of the advertisements and marketing materials say AI, they are talking about machine learning algorithms. There’s nothing wrong with that. It’s a form of AI.

Machine learning

So, what is machine learning? There is a very specific discipline to it. You get the data, which we are already collecting as part of the medical device industry, and we’ll be doing a better job of collecting it as we go on.

But, then, the most important thing that a lot of people miss out is the data as collected, most of the time, is not ready for analysis. This is the biggest challenge in cleaning up the data in what is called labeling the data.

Now if you use deep learning and neural networks, you don’t necessarily need to label the data. But for most machine learning applications, you’ll have to prepare the data, and that is also where you will have the most time, the resource, and expertise, and expenditure happen because once that happens, it’s a matter of choosing the algorithms.

And that is also a trial-and-error method. You pick a series of algorithms. Sometimes, you will run multiple algorithms on the same data, figure out what works best for you, but that takes less of time than actually working with the data itself.

Once that is done, we test the data. The idea of machine learning is that you will present the algorithm of the program, or whatever you want to call it, with new data that it hasn’t seen. So, it’s a training data, and you still want it to give you insights.

And then once it does that, there will be errors, and so, we improve that. That’s the basic flow.

Classification of machine learning types

Digging a little deeper, it’s classified this way: supervised learning, which is what most of the current applications do; unsupervised learning, which is what things like neural networks are doing with your phone, voice communications, and so on; and then there’s a concept called reinforcement learning where, when it gets it right, you give it a reward, and when it gets it wrong, you give it a punishment. There’s an error correction function, and the system corrects itself constantly over time.

If you want to think about how supervised learning happens, you want to think about it as data pairs, conceptually speaking. You will say, “This is my target value, and this is what it needs to be – when it is reaching the target value and when it is not reaching the target value.”

In terms of manufacturing, an example will be saying, “If my sensor temperature is between 50 and 60 degrees Celsius, it means that my balloons are being blown properly, for example, or my injection molding is happening at the right temperature and so on. If it is above 65 degrees Celsius, then it is not.”

So, you think of providing the machine learning algorithm with passive data on sensors. The particle count has to be this much, my ‘this much’ has to be this much, and so on. And then from there, it knows what is good, what is bad, so it can alert you. It knows what is functioning, where your process is going wrong and so on. So, that’s how supervised learning happens.

In unsupervised learning, for example machine vision, you will just feed it volumes of images or volumes of other forms of data, and what will happen is, and I’ll show you conceptually how that happens, through several layers of looking at the data, the system figures out what those features are, how to improve them. So, that’s what deep learning is.

There will be hidden layers and by hidden it means we don’t exactly see how the features are being discriminated, but eventually it knows how to recognize what is going on.

Automation and AI

So, I just wanted to quickly switch to the application side of things. It’s when you’re thinking, “So, how is AI going to improve things for me?”

Right? That’s the question you’re asking.

So, there’s all this data. There’s all this stuff. How is it going to improve?

If you look at automation today, we have automation today in most of what we do. Well, it is basically tediously programmed upfront. We know that, right?

We program it, we validate it, and we want to run this process for as long as possible without changing it because we know that if we have to change it, we have to go through all those revalidations and all that again.

AI will change that.

It is going to be a learning-based system, so it knows when to correct itself. It knows when to upgrade itself. And there is human involvement. We talked about that.

And the other thing is that today’s automation works like this – if there is even a slight problem on the line, there’s a line stoppage, and, then, human intervention is necessary,sometimes even to make minor corrections.

That will change because it knows, “Oh, a few things just fell off the floor or the tool was dulling out, so I’m just going to set aside these four guys.” It could just move them along the line, and then it’ll change the tool, and it will go forward.

So, we’re looking for those types of applications to improve. That’s what the smart factory floor is. That’s where AI will be able to improve automation for us.

AI applications

Another way to think about it is that it can solve obvious problems for us.

We can ask questions like improve my operations, show me where my scheduling is falling off, things like that; and non-obvious things – you set it on an observation path, and it’s able to tell you which machines are causing the problems to start on your floor and things like that.

AI on the floor

Yet another way to think about it is that there are problems where the AI will require human assistance. I don’t know if you’ve heard of this, but the funny thing is this: The things we struggle with, AI is able to do really well.

For example, looking at thousands upon thousands of images and being able to classify and categorize them. AI can do it like that [snaps finger].

Whereas, basic things that we do, we look at something, and we make a judgment call, “Oh. That’s going to start becoming a problem. This is going to start becoming problem.” AI has difficulties with it.

So, those are some of the different ways of thinking about it. If you did use AI for those things, there’s going to be a lot of human intervention required.

For the next 10-15 years, when you bring these things onto the floor, the left quadrant is what you’re looking at, there will be a lot of human intervention.

So, all this ‘boo hoo hoo hoo, jobs are being taken away,” that’s a big discussion. I’ve written some stuff about it. I’ve presented about it before. That’s many, many years, if that happens at all, because whatever intelligence we provide our machines or algorithms with, we still need to train them, translating our intelligence.

You know this. If you have written an assembly procedure – I come from the R&D side – you know how to build something. Getting that to other human beings is difficult enough, right? Because that’s why design to manufacturing is such a big challenge. The way we do things, they’re not able to necessarily get it across that easily.

Robotics and AI

So, if you think about another way automation be improved by AI, when you add AI to robots, you will essentially be able to change things. Algorithmic intelligence will allow the robots to function better and become more adaptable.

Industrial Robotics

I just want to touch base on these things. Everybody knows these things, but I wanted to quickly run through them: industrial robotics, we have the biggest explosion, the biggest growth going on, which is the gantries, the universal arms, the custom platforms, the platform beds.

Robotics: manufacturing and quality assurance

And then, essentially, in manufacturing, we can use robots as well. Again, I pulled a real-life application.

And, again, something I wanted to touch on, thermoforming packaging can be automated. Quite easily, actually, but it doesn’t happen that way.

So, in terms of ‘what are some of the struggles we have?’, it’s going to take many years before we are able to actually convince people to bring this on the floor.

Vision-guided robotics

Vision-guided robotics: This has been around for a couple of decades as well. The recent excitement about this is that machine vision, because of deep learning now, is more powerful, and this is where we will see some of the biggest advances.

So, some of the biggest advances that we will see first is in automation of inspection, automation of defect detection, and that will come through vision-guided robotics.

Robotics – exoskeletons

The other place where we will see vast improvements and probable applications is in exoskeletons.

One, it will improve the safety of our assembly operators; it is collaborative, not displacement. Right? It is to enhance the people who do the work on the floor, not to displace them.

The other way to think about where exoskeletons can be a big boost for your production is when you apply sensors, when you’re wearing haptic gloves and when you are doing their operations, you are now able to mathematically translate that finesse, that human ability of how to put something together, into actual machine intelligence.

Because you take the data and you will be able to automate that for larger productions. This is a big problem: We have custom stuff that has to be done.

I’m sure everybody who has done manufacturing has run into this problem: We have a couple of operators or the only people who can do a set of operations. Nobody else can do it. Right?

That’s one of our biggest challenges, so these are things where the factory’s enhanced, it’s not diminished, by bringing robotics, by bringing exoskeletons, by bringing haptics onto the floor. So that’s what I wanted to point out with this example.

The future

Where is the future headed? Researchers at University of Maryland were able to just have their robot watch – and this is a set of videos I encourage you to go on YouTube.

I watched this thing trying to cook. If you gave it the right equipment, it’s able to just watch videos and learn how to put ingredients together and make things that are probably tasteful. Right? That’s what I heard. I mean, I haven’t tasted it myself.

You see the advantage. We see cooking as one of those sort of non-engineering, very human decision-based activities, and we’re able to get the robots to do that.

So, that’s what it means for our floor because our goal in the devices industry is if I give you the device master record, you should be able to produce it without any further instruction from me. That’s where the robots will be able to help us in the future.

Human-in-the-loop

I know I’m running out of time. I just wanted to touch base on human-in-the-loop. This is a concept that promises us great research, but there’s also a little bit of a challenge with it.

Human-in-the-loop is great if we say we are going to make the ultimate decision, but, remember, you’re going to need to understand how the AI is coming at the decision that it’s asking you to approve and also to be able to do it fast, otherwise we become the bottleneck.

Explainable AI

And, then, there’s some other stuff I don’t have too much time on. I wanted to touch base on this. One of the big challenges is we still don’t know how AI makes most of its decisions, so where we want to end up tomorrow is over there, on the right-hand side.

Let me just leave you with this thought: When we needed to improve transportation, we didn’t go about automating the horse, we invented the car. We as humans, I’m not making any claim to it. The reality is new technology always scares us.

It scares us when it’s new, but, today, we don’t think of cars as scary things, right? So, it’s okay to expand ourselves. It’s okay to try new things. It’s okay to experiment with newer technologies and bring it on to the society. Thank you.

Joe Hage: I think that’s great, and it’s a great point to end on. We’ve seen it in our conversations about blockchain, on 3D, on AI, on automation. You’re absolutely right, and, at the risk of being cliché, the future is here, and we can either embrace it or we can find ourselves out of work, I think.

Srihari Yamanoor: Yes, that’s true. Somebody said, “The future with either roll with us or roll over us.” Or we can roll with it or get rolled over.

Joe Hage: Sri, thank you very much.

Srihari Yamanoor: Thank you.

Joe Hage: Solid presentation, thank you.
[Applause]