AI and Robotics in Conveyors: Enhancing Warehouse Efficiency
47:30 Share

You don't want your employees to be idle, so why would you want your robots to be idle? Zebra Technologies is addressing idle robots by bringing their latest solution to the market, which requires 30% fewer robots.

resulting in increased asset utilization and a cost-effective entry to automation for your fulfillment operations. The Zebra Symmetry Fulfillment Ecosystem is bringing you balanced utilization, more cubic capacity, and optimized pick paths so you can operate at peak performance. Want to take your fulfillment operation to the next level? Learn more at zebra.com or click the link in the show notes.

The New Warehouse podcast hosted by Kevin Lawton is your source for insights and ideas from the distribution, transportation, and logistics industry. A new episode every Monday morning brings you the latest from industry experts and thought leaders. And now, here's Kevin.

Hey guys, it's Kevin with The New Warehouse and I'm here at the MHI headquarters and I'm going to be joined by some representatives of the Conveyor Sortation Solutions Industry Group. I'm going to be joined by Zach Steck over here who is the Market Specialist for Material Handling at Pepperdell & Fuchs and also Phil Varley who is the CTO at Tenevis. And today our topic is going to be around AI and robotics

in conveyors. We're going to be talking about AI applications and the relationship between hardware. And we're also going to be talking about AI applications and their relationship with quality control, as well as some predictive maintenance as well, and what they're seeing in the industry. So Zach and Phil, thanks for joining me today. How are you both? Doing well. Thanks for having us. Absolutely. Absolutely. Good to talk about this. I think, you know,

AI, there's certainly a ton of discussion in our industry around AI, and I think every industry at this point, too, even in our daily personal lives, as well as having an impact there. And it's obviously becoming very prevalent as well, not just...

gimmick anymore in a sense. It's actually applications that are happening. But as you guys are focused on the conveyor side of things, right, which involves a lot of hardware, right? There's a relationship there between AI and hardware, especially around sensors in particular, kind of go hand in hand. Tell us a little bit about that relationship between AI itself and hardware.

Yeah, so I could kind of start there. So with AI, it's always relying on information. So the AI that we're used to is kind of relying on the internet to pull that information. But when you get into these warehouse type applications, now that information comes from industrial sensors.

So there's been trends kind of moving that direction over the last few years, and one of those things would be smart sensors. So these are standard industrial sensors that you would normally expect, except they have additional features such as additional diagnostic information, things like that.

And then there's also been more of a push to kind of get this information remotely. So kind of the leading things for that would be the IIoT side and IO-Link is a big protocol that is used pretty widely for collecting that information.

So that's kind of where we're at with that versus the normal sensors and PLC type systems. Interesting. And you're more on the hardware side, I think is my understanding. That's correct. And Phil, you're more on the AI side. So what's your perspective from that side of the relationship? It's a very interesting question that opens up the whole discussion we are supposed to have today. So classic photoelectric sensors and vision sensors, they're what we call deterministic. You set ground rules. You say, well, if...

that photoelectric sensor that photo is blocked for a certain amount of seconds you have a jam you set ground rules you set rules you follow them you know what the outcome is when we get to the ai and machine learning world this changes the the whole paradox that when when the whole industry went from relay to plc logic the the rules became more complex you have a time

portion in it. But we could never as humanity, we could never explain to a machine this is what you do just because that feels right. It's like teaching a child. You know, you can explain to a child certain things that's black, that's gray, but you can't explain morality, you can't explain intuition.

And this is where machine learning comes in. So we have smart sensors. They are your eyes and your ears. And all type photoelectric sensors are your eye, for example. And if that photoelectric sensor goes out, you realize you're blind or you don't see well.

What smart sensors allow us to do is, well, now I have a pain or I feel my eye being dry or I got something in my eye. This is a smart sensor. So I still see that I don't see well from my eye. But now I have all those feelings around my sense, around my eye. And I can say, well, I know that I'm not seeing well because. And now I can find the root cause. I can attend to that.

And that ties into AI and machine learning so much that the resolution requirement for machine learning is a lot lower than we're used to for vision systems. For vision systems, and I'm sure we'll cover that later, you have a requirement to make your decision deterministically. You set a rule and then you follow that rule. It needs to be exact.

With machine learning, you need your hardware, you need a lot of input, just like your body would. But it opens us as humanity, as an industry, up to completely different solutions and approaches that we didn't know we had in the past.

So we're going from rule-based to intuitive learning-based. So it's, as Zach said, IIoT allows us to have the data not on a controls level but on a cloud level to go from the two extremes to another. And AI and machine learning did the same thing. It went from

the cloud level from very big server requirements and computing requirements, now we're on the edge computing side. So our data security requirements and everything is met. Our data is encrypted because we don't have to move it. We don't need internal connection. It becomes an island solution. This is exactly what we want from industrial I.O.,

And this is the change that we've seen over the years. Yeah, absolutely. I think it's very interesting. And I love the analogy there to, you know, a child and trying to get them to understand those types of things. And I'm thinking about my son when you say that, and he's in the phase of, well, why, you know?

and trying to teach them that stuff. So it's a great analogy there to start to understand that. And I think even as you mentioned in there, now we understand like, oh, my eye is dry, right? In a sense, as you said there as the example. So that kind of leads into the next question there where we see AI being utilized heavily throughout different parts, right? But when we talk about something like that where...

we're understanding that there's a problem going on. Then we talk about predictive maintenance. So tell us a little bit about some of that problem detection and then the predictive maintenance and how that's leveraging AI to ultimately maximize the conveyor uptime in our systems and operations. We have all kinds of sensor inputs. We just discussed photoelectric sensors and vision, but there's also vibration, there's also temperature.

and

With normal industry sensors, you know that you have a certain temperature or you have a specific vibration, but how much do you trust them? In the aerospace industry, you have triple redundancy of two systems that you kind of look at, and if one of them doesn't work right, you have a third to know which one doesn't work. The nice thing with IoT and smart sensors, now you have that in one package. Unless you lose connection, you know what's going on with that sensor.

That allows for a nice solution for wear and tear in one package. You don't need to have a multitude. So, I mean, vibration sensors, you guys...

provide vibration sensors and those are a good primary source for wear and tear motors, correct? That's right. So yeah, I mean identifying the wear and tear becomes really important because you want to minimize downtime on these machines. So if you can kind of know when things are going to fail, you can kind of replace them before they actually do fail to reduce that downtime.

So the example that Phil's kind of talking about with vibration sensors, the way you can kind of utilize that would be to put that on maybe a motor on a conveyor belt. So as that motor begins to fail, you're going to see more vibration on the system that kind of tell you that this motor is going bad.

So that's some of the more diagnostic information that you can get with these type of sensors. And you also have information on the runtime, so how long each device has been installed in the field. So it gives you a good idea as to the life of the sensor.

And then also when you kind of go back to that dry eye scenario when we talk about photoelectric sensors, generally what they're doing is they're kind of looking at the received light strength levels back and kind of monitoring that. And if it sees a drop, then that kind of tells you these warnings are occurring or something's happening. Maybe there's dirt or dust building up on the sensor where it's not able to see its target or the reflector as well as it had in the past.

So that's kind of what we're talking about on the sensor side that can indicate early signs of failures on these machines. Interesting, yeah. And I think that's such a great thing too because like you said, we want to

minimize that downtime and maximize the uptime there, which I think is so important for any operation to be able to do that. But I guess maybe tell us a little in contrast, prior to bringing in this AI and these additional sensors to be able to do that, what did that kind of predictive maintenance look like? Was it more challenging? What was the previous state before this technology?

Zach just mentioned sensors. And as long as we stay with smart sensors and IoT and all that kind of jazz, you still have a singular unit. It's one unit providing you data and you can make decisions. Again, rule-based decisions based on the sensor. So if you have a motor and you know it's running a certain frequency, the rotor is running a certain frequency, you know that this frequency on the vibration sensor you expect.

What do you do if it doesn't?

So what if you're seeing, let's say 240 Hz being a major frequency because you're running 60 Hz, then it's a multiple. Well, now it changes. There's a couple of things you can do. You can either have rules from experience or you can have people with experience looking at your frequencies. And you notice it's the experience in both statements. Well, what if you don't have that experience? What if your system is complex enough that it can change?

VFD driven systems are the biggest problem for that because you barely can set any rules. So this is where the machine learning and AI comes in. Number one, you can bring in multiple sensors. You can bring in temperature, vibration, humidity, which a lot of customers and operations underestimate, but it's a huge issue. Humidity changes in facilities, especially in the south of the U.S.,

they can wreak havoc on all your operation and frequency analysis just because of how everything changes, especially on larger sortation systems. So now machine learning allows you to have a virtual person sitting and looking at all these inputs simultaneously. You don't have one sensor that says, hey, that frequency I didn't expect. Go take a look. Now you have a system that you can...

train by providing it feedback. It's the same as training a child or a person saying, well, you're seeing this frequency and this temperature and this is happening in the system. Load is a big thing. I have very heavy product. I've got light product.

And this is okay or this is not okay. So now you provide the feedback to the machine learning algorithm or to the AI. Not going into detail with ML and what's AI, it doesn't matter for this discussion. But now the machine learning aspect is you have all these sensors and you can A, bring them together, and B, you can set rules. For most of the stuff we're doing for predictive maintenance, it's intuitive.

The exception, because robotics is one of the topics today, the exception is there's been great algorithms even 15 years ago that I worked on where a robotic arm, you can't predict a failure of a robotic arm two weeks in advance, plus minus four hours. Not ML-based, not AI-based. It's a mechanical system. It's repetitive. You can set rules.

The rules are very strict because you have robotic arm, does always the same motion. You look at energy consumption and how long it takes to do that motion. Once you're past a certain threshold, you know that you have exactly two weeks, plus minus four hours. It was on the dot. But what if you don't have that reproducible behavior? What if you have changing loads, changing weights, changing temperatures?

That application that I was referencing was a paint shop. Your temperature and humidity is controlled. You know exactly what's happening. And it's clean. It's the exact opposite of Korea Express parcel, where you have one customer that goes through with a vacuum every day and another customer where the vacuum is kind of written off and somewhere else. There is the vacuum, yeah. Exactly. And this is the huge change from...

Old style, I just have a simple vibration input to have a multitude of smart sensors that provide me information. And now I have a model, machine learning, AI, it doesn't matter, that takes in the multitude. And from experience, just like a child, hey, it doesn't feel right. You as the father provide the feedback, yep, you're correct. And then your child learns something or you provide the feedback, nope, that's just wrong.

a feeling it's not true. And that's also where the child learns. And this is where machine learning comes in that

we as a human can provide the feedback. So, and with that we can continuously improve the process. CIP, Continuous Improvement Process, we use that for projects, we use that for engineering, and funny enough, it is the base for machine learning. So if a person understands continuous improvement processes, my discussions with these people are so much easier. And I hope this answers your question.

Yeah, absolutely. Yeah, I mean, another big advantage too is kind of reducing equipment costs. So you kind of think back to what you were saying and asking about how predictive maintenance worked before these systems existed. And at that point, there was just kind of schedules. So, okay, we have these motors, we replace them every three years, no matter what, just to make sure they don't fail and create that downtime.

So now what we're doing with these sensors to be able to monitor it, you can kind of keep track of how the motors go. And then maybe now they last five years before there's an issue. So they're getting additional cost savings from being able to use the same motor for longer periods of time. We'll be back after a quick break.

Should I be excited or afraid of AI? The answer is yes. It's both. Things are going to look so different in the next couple of years. You will not be ready for the disruptions that are going to come. I'm Jeff Nielsen, host of Digital Disruption, the go-to podcast for technology leaders. Our generation's conflict, I think, is going to be emotional, not economic. The companies that are going to win are the ones that have a couple of characteristics. Stay ahead of the curve with Digital Disruption, available now on your favorite podcast platform.

Yeah, absolutely. And I think that that's such a great thing there too, because you said like that, the routine schedule there, like, Oh, we're just going to replace it. Cause that's what we do. That's the rule that we set. But you could maybe get some longer life out of some of those things as well. And still not run into an issue where you potentially have downtime. And I think, you know,

It's a great explanation there of how you can leverage that and be able to see not only for what's your existing equipment and then also make decisions about future equipment as well, I'm sure. So very interesting there. So that's kind of focused on the hardware itself, the predictive maintenance, and I guess you could say maybe the quality.

the quality of the equipment, right, that our product and stuff is moving through. But let's talk a little bit about quality control itself, right? So I guess let's start by talking a little bit about from a quality control perspective, what are some of the challenges that you see companies facing when they're doing manual processes around quality control? And how can some of those manual processes be addressed utilizing some form of AI?

The main challenge with quality control nowadays is that it's the same as with automation. When you have a person that is not doing this every day of the week, or is his main, let's call it a hobby, because we all enjoy what we're doing, otherwise we wouldn't be, you don't know what can be done. If all you're driving is a 70-horse car,

power a car all day long, you don't know there's V10s out there with very nice engines. And it's the same when it applies to technology. Machine learning is non-intuitive. People do not understand how it works and how it doesn't. I mentioned non-deterministic. So even when we deploy

solution for a customer. The standard question is, well, why did it make this decision? And I got that engineering shulker shrug. It's like, I don't know. I can trace it. I can reproduce it. I've got the audit trail. I've got all that fancy stuff that is really helpful in 2024 industry and automation. But it's not rule-based. So the main QC blocker nowadays is

not knowing that you can automate it. So there's a quote from a long, long time ago. I'm a born lever puller. So we used to have people pushing buttons instead of industrial automation. Nowadays, we can automate a lot, especially with vision, with OCR. And the failback for OCR is

let's say five years ago still was a coding queue. You had people at laptops or at computers and keying in what the computer didn't read from optical character recognition. Well nowadays you don't. Nowadays you have models, neural networks that are specialized on vision that can code this. It's faster, it's reproducible, you have an auditory. So what I'm trying to get to is you need

The biggest blocker for quality control nowadays and to automate it is people don't realize it can be automated. If you have operations and you work day to day, you have a certain view, a certain process that is set and automated.

To break this is impossible without somebody who sees other processes. This is where PNF comes in and says, "Yep, it's a great sensor. You have that, but you can improve your uptime and everything else by using this." And this is where we come in on the ML side.

an AI side that says, "Yup, you have somebody there that looks at each and every package and says, 'Hey, the label is applied correctly or not.'" Funny enough, a correctly applied label is a tough job because, sure, you can use a vision sensor and measure whether you've got a tenth of an inch or a thousandth of an inch. Well, is it glued on or not? Funny enough, a human can determine this very well.

Measurements cannot because you can't set rules. So where does it apply on the box? Where does it apply on the GIPHY? So you have a combination of it's very hard to train people because you can't set your rules down. You have a lack of people to train. So going to the other challenges part, you have a lack of people to train.

And what do you do when you exchange the people who did that? Because you don't have rules, so you need to train the trainer. And that requires more people, more qualified people. And none of these tasks are easy, just from my experience. So now the biggest challenge is twofold. One is it's the process.

because it's not rule-based. And number two is, from our standpoint, it's a customer or the operations or the operator of the equipment not understanding, hey, this is something we can actually do. It's shocking when we go through facilities and we ask them, well, why do you still have a person standing there? Well, what do you mean? It's like,

It's a simple solution. Our bread and butter nowadays is detecting doubles on conveyors and cross-belt carriers and shoe sorters. And not only does our machine learning algorithm, does it better from a quality perspective, makes less mistakes, doesn't get tired, and doesn't get distracted by his or her phone.

And when you reproduce this, you have a talk with the customer and then they say, well, okay, what if you take this and apply it to a different problem? You know, what kind of numbers can you get? Because with photoelectric and vision sensors, we can say, well, we can detect a hairline crack of, you know, one though or two though. With machine learning, you cannot. You do a field study, you see how well you can get your model, and then you go from there.

And all of those challenges make it hard to grasp, the same as you would with a new employee. If you have a new engineer that you want to bring on board, sure, you can have an entry interview and everything else, but how well is he going to perform?

It's the same engineering shoulder shrug. Yeah, so I would agree. There's limitations to what a human worker can do when it comes to terms of inspecting. So kind of what Phil was saying, let's say there's a hairline crack, the time it would take for the human eye to pick that up or to look for it versus an AI solution, which is combining technologies can kind of get that information a lot quicker and then react to it faster.

So that's a big advantage when you get into quality inspection along these machines. Yeah, absolutely. And I think that it's, you know, that human element to it there. I mean, there's certainly some nuances I think that humans can do, right, that we've kind of touched on a little bit. But, you know, from I think a big thing that stands out to me, right, and from my experience in the industry as well, I mean,

is quality control is a it's a difficult task as you mentioned right and you need a certain mindset certain person kind of with a discerning eye I guess you could say to be focused on quality control and be able to catch those kinds of things whereas now you see you know that level of detail attention to detail in that sense is hard to come by right it's not like you can just say oh we have somebody that's a

there's not enough orders to pick right now, let's just throw them in quality control, right? There's a lot of training, understanding, repetitions that need to go into that to do that. So I think bringing in that AI and machine learning, as you said, certainly can tackle some of that as well. So how do you see, I guess, on that human element involved with quality control? I think you touched on it a little bit here, Phil, but how do you see AI kind of

evolving and how does it impacting the human in that role? Are we seeing that it's replacing them, it's assisting them, making it easier for them? Where does that relationship kind of stand right now from a quality perspective?

That's one of the things. So with AI, there's always a fear of it replacing people, especially in jobs. So that's not really the idea that we want to work towards here. So assisting is the better term. I mean, replacing is well, but not replacing them in the position, but allowing them to kind of focus their efforts onto things that are better suited for a human's capabilities. So making decisions, things like that, doing some more diagnostics, let's say in the quality department,

If this AI system flags a part, then the human operator can kind of go over and see if, okay, maybe this part's salvageable or it has to be scrapped so that they can kind of make decisions that way.

So it's really just kind of changing the role of the human operator in conjunction with the AI system. And I mean your statement is absolutely valid. You guys as PNF don't want to replace people. We have a bit of a different view. And it's not because we were, you know, misanthropic. No, but let's talk history here a little bit. Quality control, you have a person checking a production run. I'm taking something from my family history, heavy steel.

You have an engineer that works at the mill and he takes samples. You have a second engineer in the second shift and a third one in the third, and you have somebody to cover vacations. You have multiple people executing the same task with different quality levels, with different approaches, different processes for themselves. That was 50, 60 years ago.

You had more people than jobs. Then it kind of shifted and fast forward, very fast forward to 2024, we have a huge lack of maintenance personnel. So you used to have multiple people to do the QC. Now you don't even have enough people to do a portion of this. So what do you do? You leverage the people who are qualified to do quality control

And that applies to different other problems, not just quality control, but let's stick with that as an example. And you use these people to train machine learning algorithms to say, this is good, this is not good, this is maybe good.

with different inputs. So you're leveraging those people to actually execute the tasks you need to execute. If we wouldn't be automating it, we wouldn't be doing these tasks because we don't have enough people. The same with predictive maintenance. We don't have the people nowadays in the U.S. this year to do the maintenance we need to do. It's a huge problem. I don't see it as replacing...

I'm seeing it as filling roles that need to be filled.

And because machine learning algorithms need to be trained as well. Who trains them? Humans need to train them with experience. This is how it helps the human inspectors. And from my experience over the last 10 years, we went from assisting systems, so automation level two and three, so where a system tells you, hey, I recommend this action, and now a human takes this action,

Nowadays, most systems we roll out are level four automation. You're not aware they're there. You don't see them. You don't feel them. They don't require human input. They just work. We have facilities where operations and maintenance is surprised when we call them up and, hey, we want to do maintenance on this system. And they say, what system? They're not even aware it exists.

For predictive maintenance and maintenance support, absolutely. You need systems that either tell you to do something or make recommendations. When it comes to QC and automation, we are already on level four automated systems. We're already in the realm of machine learning is better, faster, more reliable, and same as IIoT sensors,

I have my audit trail. I have my big data approach. I can always go back and if I have all the data I can at least reproduce, not understand, but reproduce why AI took that decision or made that decision that it did. We don't want to replace people, but right now we're not in the state that we need to worry about it because we don't have the people that we are replacing right now anyway.

Yeah, absolutely. And I think we see that not just in the quality side or the maintenance side, but just in general, right, in our industry. Like the people doing that work on the floor, right, either they don't want to do that kind of work, right, or it's just hard to find them in certain areas because it's become such a competitive market for those warehouse operations workers. And I think it's very interesting there, too, Zach, you mentioned, you know, kind of the

freeing up some of those people in a sense, right? Where we talk about quality control as the example, right? You have somebody

packages, for example, right, to see, like you said earlier, feel like if the label is in the correct positioning or something like that. But, you know, you think about that and, well, how many packages is that person checking that are just perfectly fine, right? And they could be doing something else with that time. Sure. So I think, you know, bringing that in and having that as that, that is,

assist right and reducing the number of people that need to to check uh and only focus on when there is actually an issue and do some corrective action i think is a a great approach to to take there um and being able to do that um so you know with this analysis right and all of this ai and machine learning going on obviously uh we're in any different aspect we're uh

collecting tons of data, right? And then it's being analyzed by the AI going on. So tell us a little bit about how does that make operations more efficient? How does it speed up decision making? And how do we get to kind of see some of these trends a little quicker than maybe we would in the past when we have to kind of decipher and analyze some of that data manually or on our own as humans?

Like you kind of mentioned, there's a lot of things that are pretty complicated within these warehouse solutions, conveyors, systems. So technology's kind of come along

to kind of help with that. So even from the hardware standpoint, some more advanced technologies such as like the LiDAR systems and 3D vision have kind of made this possible because now you're getting all this data that you need. So it's a lot of data, but the thing with AI too is that you want to be able to access that information quickly. So one thing that a lot of manufacturers are doing is so we have these very sophisticated sensor systems

But there's also application specific software that condenses the data that's coming out of it to kind of gear it more towards a specific purpose, which could then be better analyzed by the AI system. So an example with that would be for like systems for like conveyor belt utilization, just to make sure that the system is running as efficient as possible and they're running as many boxes through as possible.

So you can have kind of some softwares that will kind of look at that give you that percentage value a flow with maybe some height information and Then the AI system can kind of process that a lot faster than these heavy data type products. Interesting. Yeah

And I'm going back to the experience and how it used to be because we can learn a lot from that. But I think all of us still know people who went into a plant or facility and listened to it and knew, hey, it just sounds off, sounds wrong.

And this is where we talked about lack of personnel. Historically, there is an engineering speciality for that. That's digital signal processing. The problem is that that's been on a decline. So we have very few electrical engineers, even less well-qualified ones, and DSP is a very tough field to study.

So we have less people that can look at data and know how to analyze it from a mathematical perspective. So now you have that multitude of data, what do you do with it? Sure, you can feed AI models with it, and this is exactly what you need to do because the amount of people that over the years I've worked with that can see through data and analyze data and say, "This is what we need to look at,"

They're very smart people. I was impressed by what they see by just looking at certain things, see a structure in something hidden. But there's no way nowadays with as many facilities as we have, as much equipment as we have, and with those engineering specialties,

On a decline, it doesn't match. The scissor just opens up here. And this is where it comes in that we have huge amounts of data. How do we analyze it? And we as an industry can start training people. So you're talking about sending somebody to schooling and then giving him experience.

That's a great long-term goal and we as a company support this, we execute this, but the problem is it's a very long-term investment.

But that's on the philosophical company side. On the product company side, we have solutions for that because we also see that people with experience are very good at going to a plant they don't know, to a facility they do not know, and see, hey, okay, what data do you have? Show me. And then we can talk about what should we be doing with that data.

So it's a changing field completely because of how the schooling changed over the last recent years and how our industry grew over the last recent years. We've had ridiculous growth. I'm very happy about it personally.

But the problem is that we're putting high-tech systems into the field and partially have nobody to support those. And we need to do something about it. Otherwise, we'll stop putting in systems. So we need to automate the automation at this point.

Yeah, I think that's a great point there because I see that too, right? And to your point, I mean, it's incredibly exciting to see the technology and stuff that's coming into our space and what we're able to do now with these operations and put in place. But in the same regard, you know, you don't necessarily see the technology

the people that have traditionally worked in these operations and facilities be up to speed on these technologies, right? And understand, you know, how to leverage them, how to deal with that data or how to even do, going back to what you said earlier in the conversation, deal with some of that maintenance too as well around those things. So there's certainly a gap there to do that. But, you know, with that data being generated and all these new solutions coming out that are creating

even more data when you put them in, right? You need to be able to take some action and do something with that. So I think AI can kind of

help digest some of that and make it a little more digestible for people to be able to take action, right? Which is the point of data at the end of the day, right? Is to be able to take some type of action on it and improve overall. So if we look here, I mean, throughout this discussion, we've talked about kind of some of the AI applications, relationships around hardware, how it can be leveraged and some of the things that are happening now in operations and can be done in operations now.

For people that are considering starting to implement further AI-driven practices, whether it's around that quality control we talked about or other processes in their operations, what are some of the key considerations and how should they go about their thinking of approaching that? What's the best way to start and what should they really consider as they start to approach these solutions?

consider getting somebody in with a different viewpoint. When you want to buy a car, you don't look at what you have in your own garage. You go to the dealer to get an idea, and then he starts asking you questions and trying to upsell, obviously, but hey. And here it's the same thing. AI and machine learning is such an abstract concept for humans right now that...

The QC we're doing right now, stay with the topic we had, that wasn't customer driven. We have a huge deployment of quality control, machine learning based, island systems. They don't talk to the cloud. They're all on their own little level on the industrial network. They don't need to communicate. There's no concerns. But that didn't come from the customer. We had a project that we executed with a customer

we saw that is a problem. It's a problem in the process and there wasn't a solution. So what we said is, well, we have that technology that we've been, let's call it, playing around with for a couple of years.

And it would fit that problem perfectly. And it did. We talked about children and learning. So the first deployment we had, I compared to about a four-year-old. Attention span and quality. We had a detection rate of 60% to 70% and a false positive of one, which for the customer was good enough. Nowadays, we're well over 90% detection rate and a false positive of way lower than 1%.

So the evolution steps are huge.

but the customer wasn't aware that this is a problem that can be solved. The customer's solution for this was, again, a bone lever puller. It was a person with a button that had a machine that would divert the defective product. Well, there's a couple issues with that. That machine to divert created more defects by itself. That person would get tired. That person would get distracted and would make wrong decisions.

So even our first throw at a machine learning algorithm already improved on that human. And then we took it a couple levels higher. So what I'm saying is from...

If you want to improve your process, maybe what you're thinking about isn't even the biggest hitter. If you did your Fishbone analysis and your 3Sigma and you know how your defects combine into your total defect number, great. That's a good start. And then we look at the largest hitters, not at what the customer thinks can be solved because maybe you can solve something different.

A lot of integrated motors with VFDs have temperature probes, have all kinds of power draw. That's a lot of data that already can be utilized, and very few people do. Does it make sense to add a vibration sensor? Absolutely. It's a very determined solution for a specific problem. It is not the solution. The solution is most of the time already present.

There's a saying about woods and trees. The way to approach this is, okay, I think I have this problem. Then finding somebody who is in that field of solving problems. The way it usually turns out is there's going to be a proposal with a system with all kinds of sensors, with a computing approach. We prefer edge computing because it gives the customer the power to touch something and the data being present.

And then training the neural networks or train the machine learning approach. The first step is get somebody involved that knows how to automate the process because those are not fixed solutions. You can't open a drawer and there it is. It's the same as training a child. You need somebody and now let's go solve the problem. Let's go build that RC plane.

So this is a very long answer to your question. Sorry. Basically, find an expert. Yes.

It's the same as with controls. The customer does not start building their own control solution. There are some customers who have groups for doing this, absolutely. But it's a dedicated group. And there's very few customers who have that because with companies like PNF, you provide solutions for a various amount of customers in different fields. You learn a lot. So the additional value of PNF is that

that accrued over the history of the company, and we talked about P&F started by a couple of guys, they came with a lot of knowledge from the original company. And it's the same with machine learning. It's just a completely different field. It does not overlap the industrial automation. We as a company are just...

blessed that we do both. This is how we grew, but your AI and machine learning solutions are a specialized field. You need somebody that knows how the training process goes to understand, okay, what kind of inputs can I use? That brings in vibration sensors, it brings in temperature, it brings in all kinds of primary sensors to then understand, okay, this is a problem I can solve. Yeah, I

Yeah, absolutely. And I think it's great, great recommendation there and insight on what to consider if you're starting to look at these and, you know, bring in somebody that knows what they're doing. They've been through this. They have the experience to do that. Zach, what would you say for key considerations? Yeah, I would say another key consideration, again, coming from more of a hardware standpoint, is being able to incorporate AI into existing systems.

so typically these facilities already have a lot of things already installed so whether it's the conveyor systems robots sensors so having that ability and flexibility with ai to be able to integrate the already existing components and then add this new feature is a big advantage but you definitely want to consider that when kind of going down that route to make sure that what you have in order to make sure it's as

cost-friendly and as smooth as possible when you kind of make this switch over is in integratable into that system. All right, awesome. And I think there's very good insights there, very good considerations to take when people want to start to approach this, start to bring this into their operation and maybe are not quite sure where to start or do or what to do to get started on that road. So it's a very insightful conversation here. I really appreciate you both

joining me here to talk about this topic and help to teach some of the listeners out there in the audience on how to go about this, what's possible, what's even out there at this time. I know, Phil, you mentioned that some people don't even realize that they can do things and solve these problems that they're having. So very interesting to talk about this and create more awareness about this. I really appreciate the conversation from you both.

And if people want to learn more about the conveyor sortation solutions group, you can find that at mhi.com. And we'll link to that in the show notes as well, as well as links to get in touch with both Phil and Zach as well. So appreciate you guys sharing your time with me and insights as well. You've been listening to The New Warehouse Podcast with Kevin Lawton. Subscribe and check us out online at thenewwarehouse.com.

Enjoyed this episode? Make sure you are subscribed to the podcast and for more content from The New Warehouse, find us on LinkedIn and YouTube. Links to subscribe can be found in the show notes and for everything The New Warehouse, head to thenewwarehouse.com.