Through the Line: Packaging and Processing
This podcast explores innovations and information across the packaging and processing landscape, from topics impacting consumer packaged goods and healthcare packaging, to the latest technologies in food processing operations. Join us for the latest insights, trends, and strategies shaping packaging and processing today.
For Mundo Expo Pack content go to: https://vocesdesdelalinea.buzzsprout.com
Content hosted by Packaging World, ProFood World, and Healthcare Packaging.
Through the Line: Packaging and Processing
Lean Principles in the Machine Learning Era: ProFood World
Use Left/Right to seek, Home/End to jump to start or end. Hold shift to jump forward or backward.
What happens when decades-old lean manufacturing principles collide with today's machine learning push in food and beverage plants?
Industry experts break down why data standardization, not model-building, is the real bottleneck to AI-driven waste reduction, from siloed IT and OT departments to inconsistent naming conventions across plants. The piece traces real-world wins in predictive maintenance and packaging-line bottleneck diagnosis, while making the case that scaling these gains enterprise-wide starts with people and governance, not more technology.
This is an AI-generated episode. Read the full featured article on ProFood World.
Welcome to Through the Line, the podcast exploring innovations and information across the packaging and processing landscape. From topics impacting consumer packaged goods and healthcare packaging to the latest technologies in food processing operations.
SPEAKER_01Hi, I'm Casey Flanagan, Associate Editor with ProFood World. This AI-generated podcast episode covers how food and beverage manufacturers are applying lean principles to machine learning and data standardization. It explores how these tools root out waste in packaging and predictive maintenance while bridging the gap between IT and OT silos. We'll also discuss the importance of building a reusable data infrastructure to successfully scale AI initiatives across the enterprise.
SPEAKER_03Today, we're looking at a pretty harsh reality in the processing and packaging industry.
SPEAKER_02Yeah, it's one a lot of people don't really want to admit.
SPEAKER_03Because you can spend millions of dollars upgrading a facility. You get state-of-the-art robotics, high-speed conveyors, the most advanced sensors on the market.
SPEAKER_02All the shiny new toys.
SPEAKER_03Exactly. But the moment you actually try to implement artificial intelligence to optimize that line, everything just grinds to a halt.
SPEAKER_02It's incredibly frustrating for plant managers.
SPEAKER_03Because the machines are generating terabytes of data, but none of it connects. So here is the one-sentence summary for today. The future of artificial intelligence in manufacturing doesn't rely on algorithms, it relies entirely on standardizing your data first.
SPEAKER_02And the overarching issue we're seeing right now is just this massive disconnect between expectation and reality. AI consistently grabs all the industry headlines, right? It promises autonomous operations and these massive efficiency games. But the real kind of unglamorous groundwork in the food and beverage industry right now, it involves standardizing data across batch processing records, lab databases, and the physical production lines themselves.
SPEAKER_03So the not-so-fun stuff.
SPEAKER_02Exactly.
SPEAKER_03Yeah.
SPEAKER_02And if that data isn't standardized and accessible, the most advanced machine learning algorithms in the world can't generate a single actionable insight. They just process garbage in and spit garbage out.
SPEAKER_03Yeah, and the stakes for getting this right are actually incredibly high. According to the 2023 State of Lean Manufacturing Report, only 10 to 15% of U.S. companies systematically use lean principles to reap competitive and financial benefits.
SPEAKER_02That is a wild statistic, only 10 to 15%.
SPEAKER_03So our mission in this deep dive is to understand how applying lean principles specifically to your data and not just your physical processes is the fundamental key to unlocking AI's true potential in packaging and processing.
SPEAKER_02And to understand where machine learning is going in the packaging sector, we really have to look backward first.
SPEAKER_03Aaron Powell, the legacy of automation.
SPEAKER_02Exactly. We need to look at where that first wave of modernization sort of fell short regarding data strategy.
SPEAKER_03Aaron Powell Which brings up John Oskin. He's the senior vice president at Smart Sites, and he provides some great insight on this in our source material.
SPEAKER_02Aaron Powell Yeah, his points are spot on.
SPEAKER_03He points out that most food and beverage producers, they've invested heavily in automation over the past, say, 15 to 20 years. They installed thousands of programmable logic controllers, modern sensors, all of that.
SPEAKER_02But there was a catch.
SPEAKER_03A huge catch. He notes that data standardization was rarely part of the original charter. Plant managers were basically just focused on getting the machines running faster, but there was no master plan for how all the generated data would actually be unified later on.
SPEAKER_02And that lack of a master plan, it created the fragmented systems we have to deal with today. David Arians, the founder of the ITOT Insider, he traces this back to how we traditionally think about waste.
SPEAKER_03Okay, physical waste, right?
SPEAKER_02Yeah. Historically, methodologies like Lean and Six Sigma gave companies the rigorous discipline needed to eliminate physical waste on the plant floor.
SPEAKER_03So reducing scrap material, cutting down on unnecessary physical movement, controlling process variability, things like that.
SPEAKER_02Aaron Powell Exactly. But Arians explains that the industry now has to apply that exact same discipline to its data. Companies have to eliminate waste in how information is found, cleaned, and contextualized.
SPEAKER_03So treating disorganized data as actual manufacturing waste.
SPEAKER_02Yes. Because the ultimate goal is to build an infrastructure where a data team doesn't have to start from absolute zero every single time they want to launch a new machine learning use case.
SPEAKER_03Which makes a lot of sense. And we're actually seeing the shift from physical lean to data lean impact, how entirely new facilities are built.
SPEAKER_02Oh, absolutely.
SPEAKER_03Yeah, in a recent Pro Food World article focused on standardizing production metrics, Bob Rice, he's the vice president of engineering at Control Station Inc., he highlighted a massive shift in new construction.
SPEAKER_02What's he seeing there?
SPEAKER_03He noted that 20 years ago, greenfield projects were solely focused on getting the equipment installed and running. But today, greenfield food plants are taking a data first approach.
SPEAKER_02That is a huge paradigm shift.
SPEAKER_03It really is. They require strict operation standards and functional analytics well before the first physical production run even happens. They're actually mapping the data architecture at the exact same time they pour the concrete.
SPEAKER_02That's incredible. But obviously, most of you listening aren't building brand new greenfield plants from scratch.
SPEAKER_03You're managing legacy systems.
SPEAKER_02Yeah, it's like building a state-of-the-art automated commercial kitchen, but realizing none of your new appliances plug into the same type of electrical outlet.
SPEAKER_03Oh, that's a great analogy.
SPEAKER_02Think about it. The blender has a three-prong plug, the oven needs a 220-volt connection, and the refrigerator only runs on some proprietary battery.
SPEAKER_03Sounds like a nightmare. So how do existing plants even begin to retrofit their operations without what the industry calls a full manufacturing ontology?
SPEAKER_02That machine readable model of all their equipment and process.
SPEAKER_03Exactly. If you don't have that, where do you start?
SPEAKER_02That kitchen analogy hits right at the core of the problem. A full manufacturing ontology explicitly defines how every piece of equipment relates to a process, how those processes consume materials, and how batches follow specific recipes.
SPEAKER_03But most existing facilities don't have that perfect universal framework.
SPEAKER_02No, they don't. Yeah. So they can't just overhaul their entire infrastructure overnight. The plants that are actually succeeding right now, they're the ones starting very small. They're securing immediate, measurable wins just to prove the concept.
SPEAKER_03Okay, let's look at what starting small actually looks like in practice. Michael Warder, who is the senior vice president and chief information officer at Ruiz Foods, he shared a really compelling example at the 2024 IFT first conference.
SPEAKER_02Yeah, Ruiz Foods is a massive frozen food producer.
SPEAKER_03And they initiated a data standardization project specifically isolated to their research and development business unit, just RD. The goal was simply to move the team away from disparate systems and manual spreadsheets into one unified database.
SPEAKER_02And Warter noted that just taking that one initial step of standardizing RD data, it immediately yielded regulatory compliance and track and trace benefits.
SPEAKER_03Oh wow, just from one department.
SPEAKER_02Yeah. Suddenly tracing an ingredient from a supplier through the testing phase didn't require opening dozens of different spreadsheets.
SPEAKER_03That sounded like a massive relief for the team.
SPEAKER_02It was. And by proving the concrete value in one department, Warter was able to show the Ruiz Foods Board the direct connection between applying lean principles to data, reducing waste, and actually building the foundation for future AI capabilities.
SPEAKER_03That makes sense. Let's drill down even further, though, from a departmental level to a specific packaging line case study.
SPEAKER_02Okay, let's do it.
SPEAKER_03There was a recent webinar titled Optimizing Production with AI and Machine Learning, featuring Mark Bertrand at Smart Sites. A customer brought them in to diagnose a severe, ongoing bottleneck on a packaging line.
SPEAKER_02Those are always tricky to pin down.
SPEAKER_03Very. The operators were tracking different key performance indicators, or KPIs. Specifically, they were looking at mean time between failure's MTBF versus the effective rate.
SPEAKER_02And just for context for the listener, effective rate is calculated by multiplying equipment availability by the average rate for the line.
SPEAKER_03Giving you a comprehensive units per minute metric.
SPEAKER_02Exactly. So the physical equipment in this scenario included a bundler, a wrapping unit, and a tray packer. Using SmartSite's ABLE technology, they conducted a root cause analysis across all three machines.
SPEAKER_03What did they find?
SPEAKER_02The AI model identified the tray packer as the highest impact machine based on the potential root causes for failure.
SPEAKER_03Okay, so the tray packer is the problem.
SPEAKER_02But wait. In parallel, they ran a prescriptive analytics model on the entire packaging line, focusing entirely on that effective rate KPI we just talked about.
SPEAKER_03Oh, interesting.
SPEAKER_02And that second machine learning model revealed something totally different. It showed that the operational focus actually needed to be on the bundler, not the tray packer.
SPEAKER_03Wait, really?
SPEAKER_02Yeah. And by shifting focus to the bundler, the ML modeling enabled the plant to increase over speed capacity and the overall rate on both machines, which successfully resolved the bottleneck.
SPEAKER_03Hold on, I'm stuck on something here. If one artificial intelligence model points a floor manager to the tray packer and the other model points them to the bundler, how is that manager supposed to know which one isn't lying to them?
SPEAKER_02That's the million-dollar question.
SPEAKER_03Because don't conflicting AI insights just create more chaos on the floor. Mark Bertrand noted that both algorithms were technically correct, but how does that actually work in reality?
SPEAKER_02It works because an algorithm is only answering the exact mathematical question you ask it.
SPEAKER_03Oh, okay.
SPEAKER_02The root cause analysis model was likely optimizing for machine uptime and identifying microstops. So if the tray packer jams for 30 seconds, 10 times a schist, the algorithm flags it as the primary source of failure. Because the mean time between failures is incredibly low.
SPEAKER_03Because it stops constantly.
SPEAKER_02Exactly. But the plant manager's actual goal isn't just to stop the tray packer from jamming, right? Their goal is to maximize overall line throughput.
SPEAKER_03Which is where the effective rate comes in.
SPEAKER_02Yes. That is why the prescriptive analytics model, guided by the effective rate KPI, correctly identified the bundler. Even though the bundler might fail less often, it was the actual constraint limiting the total units produced per minute for the entire line.
SPEAKER_03Aaron Powell So the context of the KPI matters entirely.
SPEAKER_02Entirely. The AI has absolutely no idea what your broader business goals are unless you select the appropriate metric to guide its analysis.
SPEAKER_03Okay, so the technology is clearly capable of fixing complex packaging bottlenecks when you guide it with the right KPI. So naturally, why hasn't every processing plant seamlessly scaled this across all their lines?
SPEAKER_02Yeah, you'd think they would.
SPEAKER_03Marcus Gerster, the founder and CEO of Mont Blanc, he argues that the roadblock isn't the math. It's about operational adoption.
SPEAKER_02I completely agree with him.
SPEAKER_03He says that applying lean principles in the machine learning era means embedding these insights directly into daily workflows.
SPEAKER_02Gerster is entirely right on this. You can build the most elegant machine learning model in the world, but if it doesn't connect to the morning production meetings or the daily maintenance routines, it fails.
SPEAKER_03Just a shiny dashboard nobody uses.
SPEAKER_02Exactly. If the floor operators don't trust the data, or literally don't know how to use the dashboard, the technical brilliance of the model generates zero sustained impact.
SPEAKER_03And David Arians points out that this lack of operational adoption usually stems from a massive organizational divide. The classic battle between IT and OT.
SPEAKER_02Oh, yeah. Information technology and operational technology.
SPEAKER_03Arians provides a very sharp analysis of this dynamic. IT teams and OT teams, they operate in complete silos in most manufacturing companies.
SPEAKER_02You really do.
SPEAKER_03They have different reporting lines, different budgets, and most critically, they speak entirely different languages regarding risk and success.
SPEAKER_02Yeah, if you think about it, your IT department is typically focused on enterprise data security, cloud architecture, and preventing network breaches.
SPEAKER_03While the OT department is focused on physical plant safety, hitting daily production quotas, and keeping the conveyor belts physically moving.
SPEAKER_02OT thinks in terms of milliseconds and shifts.
SPEAKER_03That's a huge disconnect.
SPEAKER_02It is. And Arians warns that if those two worlds don't cooperate and align their goals, no amount of advanced technology can save a company's data strategy.
SPEAKER_03And Arians actually outlines two sequential data bottlenecks that occur even after you get IT and OT to sit at the same table.
SPEAKER_02The first bottleneck being the sheer lack of reusable infrastructure.
SPEAKER_03He estimates that 60 to 80% of project time is wasted on basic plumbing work.
SPEAKER_02We call it plumbing work because it is tedious manual labor. When a company decides to launch an AI project, data scientists spend weeks just finding, extracting, aligning, and cleaning data from entirely disconnected sources.
SPEAKER_03Which is not what you want your highly paid data scientists doing.
SPEAKER_02Definitely not. A typical plant has historians tracking sensor data, manufacturing execution systems tracking orders, computerized maintenance management systems, logging repairs, lab information systems checking quality, and ERP platforms tracking finances.
SPEAKER_03And none of these platforms were natively designed to speak to each other?
SPEAKER_02No, not at all. A historian tracks a valve pressure in real time, while an ERP tracked the cost of the batch completed last Tuesday.
SPEAKER_03So when a company moves to their second AI project, they almost always have to repeat that exact same grueling plumbing process from scratch.
SPEAKER_02Because they didn't build a permanent bridge between the systems the first time.
SPEAKER_03And that leads right into the second bottleneck Arians identifies, which is context and governance.
SPEAKER_02This one is huge for scaling.
SPEAKER_03Because even if you manage to wire all those disparate data sources together, you still hit a wall if every individual plant defines its assets differently.
SPEAKER_02Oh, the naming conventions.
SPEAKER_03Yeah. If plant A in Ohio logs a machine as conveyor one and plant B in Texas logs the exact same machine as packaging Belt Alpha, the central AI model can't analyze them together.
SPEAKER_02It just sees two completely different things. If you lack agreed-upon naming conventions and data structures across your enterprise, you haven't built a reusable infrastructure.
SPEAKER_03So how do you fix that?
SPEAKER_02Scaling machine learning requires standardized governance. Arians suggest grounding these standards in established frameworks like ISA 95.
SPEAKER_03ISA 95.
SPEAKER_02Yeah. It's essentially an international standard for developing an automated interface between enterprise and control systems. It acts as a universal dictionary.
SPEAKER_03Okay, so it ensures that data context is portable across different manufacturing sites.
SPEAKER_02Exactly. It doesn't matter if you start your first project analyzing historian data or lab data. What matters is that you are building infrastructure that compounds in value rather than building isolated point-to-point connections that just can't be scaled later.
SPEAKER_03I want to distill this with a thought for you to consider regarding your own facilities. If your IT department and your floor operators speak completely different languages about what constitutes a win, are you really building a foundation for the future?
SPEAKER_02Probably not.
SPEAKER_03Or as Arians bluntly asks, are you just spending millions of dollars to build neater silos?
SPEAKER_02Neater silos. That hurts, but it's often true.
SPEAKER_03Yeah.
SPEAKER_02But when you finally do bridge that gap, overcome the silos, and finish the hard plumbing work, there is a very specific immediate payoff waiting for you.
SPEAKER_03What's that?
SPEAKER_02For packaging and processing facilities, the most immediate measurable reward is predictive maintenance.
SPEAKER_03Ah, okay. And the industry data backs that up heavily. Leva recently released a report titled 2026: The State of AI in Consumer Goods. They surveyed 150 consumer packaged goods, senior quality and IT leaders.
SPEAKER_02And what were the results?
SPEAKER_0352% of those leaders explicitly stated that predictive analytics is their top AI priority. Furthermore, nine out of 10 respondents say their companies are actively using AI or are currently conducting pilots.
SPEAKER_02That's massive adoption. And Marcus Gerster explains exactly why predictive maintenance is the logical first step.
SPEAKER_03Why is it the low-hanging fruit?
SPEAKER_02Because equipment failure events are concrete and highly measurable. A packaging line motor is either running or it has failed. It's a binary state.
SPEAKER_03It's not subjective.
SPEAKER_02Exactly. That binary nature makes labeling the data incredibly easy when you're trying to train a machine learning model. Contrast that with trying to optimize a complex process, like, say, product flavor profiles or texture.
SPEAKER_03Oh, yeah. That would require deep, messy integration with subjective lab results and variable raw ingredients.
SPEAKER_02Exactly. Furthermore, the return on investment for maintenance is simple to crowdfy. You literally just measure the reduction in unplanned downtime and the money saved on emergency spare parts.
SPEAKER_03John Oskin adds another practical layer to this too. He notes that production assets are already producing the exact type of data AI needs.
SPEAKER_02The data is just sitting there.
SPEAKER_03Yeah. Modern packaging lines are covered in sensors monitoring vibration, temperature, electrical current, pressure flow. This structured high-frequency data is the ideal fuel for machine learning algorithms.
SPEAKER_02And companies like smart sites use this continuous data flow to automatically trigger maintenance work orders in the system or issue alarms when scata anomalies are detected.
SPEAKER_03And for context, a SCADA system supervisory control and data acquisition is what operators use to monitor the physical plant processes.
SPEAKER_02So when the AI detects a vibration anomaly in the SCADA data, it flags the issue before the bearing actually shatters.
SPEAKER_03Which saves incredible amounts of time and money.
SPEAKER_02It does. But while those isolated successes are well documented, David Arians offers a pretty stark warning about scaling predictive maintenance.
SPEAKER_03What's the catch?
SPEAKER_02A proof of concept that brilliantly predicts bearing failures on one specific packaging line can fail completely when you roll it out to another line.
SPEAKER_03Why? If it's the exact same machine. Wow.
SPEAKER_02Yeah. It can fail if an operator manually tweaks a machine set point to handle a different packaging material. Or, most commonly, the predictive model simply degrades over time because nobody in the organization actually owns its ongoing maintenance.
SPEAKER_03Wow, so Arians emphasizes that scaling a predictive model across an entire fleet of assets demands the exact same reusable data infrastructure and governed context models we discussed earlier.
SPEAKER_02You just can't escape the plumbing work.
SPEAKER_03No, you can't. But despite these massive hurdles, John Austin notes that corporate managers are fundamentally changing their stance. Due to intense cost pressures and shrinking margins in the food and beverage industry, leaders are finally taking a leap of faith on AI.
SPEAKER_02It's moved from a nebulous concept to a required tool for remaining competitive.
SPEAKER_03But taking that leap of faith requires an honest understanding that artificial intelligence is not a plug-and-play solution.
SPEAKER_02Definitely not. You can't purchase an algorithm, point it at a disorganized database, and expect it to magically resolve your production bottlenecks.
SPEAKER_03No. Let's bring all these threads together so you can apply them to your own processing and packaging facilities. Treating AI as a quick fix just won't work. True innovation requires returning to the basics and applying traditional lean principles to your data.
SPEAKER_02Absolutely.
SPEAKER_03You have to be willing to do the tedious plumbing work to standardize your inputs across your programmable logic controllers, your manufacturing execution systems, and your enterprise resource planning software.
SPEAKER_02And you have to tear down the operational walls between your IT and OT departments. They need to share the same goals, the same timelines, and the same vocabulary regarding risk.
SPEAKER_03Yes, and you must start with small, highly focused, KPI-driven projects like resolving that specific bundle or bottleneck we talked about before you ever attempt to scale a solution across your entire enterprise.
SPEAKER_02The foundational message from all the industry experts we review today is remarkably unified. Don't focus on the algorithm first, focus on the infrastructure.
SPEAKER_03Get the plumbing right.
SPEAKER_02Standardize the data, govern the context with frameworks like ISA 95, and ensure the operators on the plant floor actually trust and integrate the insights into their daily workflows.
SPEAKER_03Exactly. I want to leave you with a final question to mull over as you walk your plant floor this week. We established earlier that 60 to 80 percent of a plant's data project time is currently eaten up by basic plumbing just finding and cleaning disconnected data.
SPEAKER_02That's a lot of wasted potential.
SPEAKER_03It is. Think about the sheer scale of that wasted time across your entire organization. What revolutionary process improvements are we leaving on the table simply because our machines and our departments don't speak the same language?
SPEAKER_02It's a great question.
SPEAKER_03What if your next major packaging innovation doesn't come from purchasing a new piece of hardware, but from finally organizing the data you've been ignoring for a decade?
SPEAKER_02Something to think about.
SPEAKER_03This has been your custom deep dive into your requested sources. Keep optimizing.
SPEAKER_00Thank you for listening to Through the Line Packaging and Processing. You can listen to more episodes on all streaming platforms. Be sure to visit us at packworld.com, profoodworld.com, and healthcarepackaging.com for more packaging and processing news. This podcast was edited by Bree Guns.
