Best of LinkedIn: Smart Manufacturing CW 21/ 22

00:00:00: This episode is provided by Thomas Allgaier and Frennus, based on the most relevant LinkedIn post about smart manufacturing in calendar weeks twenty one.

00:00:08: And twenty two.

00:00:10: Frenness is a B to B market research company that supports smart manufacturing providers with building feature-by-feature competitive intelligence that shows exactly how their product stacks up against the competition.

00:00:22: you can find more info in the description.

00:00:24: we really need that kind of intelligence right now because this space has just.

00:00:28: It's moving so fast.

00:00:29: Oh, totally!

00:00:30: I mean welcome to our latest deep dive.

00:00:32: everyone.

00:00:33: today we are basically cutting entirely through the vendor hype.

00:00:36: right.

00:00:36: no fluff today

00:00:37: none.

00:00:38: We're looking strictly at what professionals on the ground or sharing right now about smart manufacturing.

00:00:43: Yeah extracting the core themes from the actual conversations happening across LinkedIn.

00:00:48: exactly.

00:00:49: So whether you know currently standing on a factory floor trying to commission some stubborn machine, or if you're in the boardroom mapping out at completely new facility we are unpacking actual mechanics of what's working and totally failing.

00:01:01: And honestly there is really only one place where can start

00:01:05: The biggest buzzword.

00:01:06: all right?

00:01:07: Yep

00:01:07: industrial AI and agentic manufacturing.

00:01:11: But before getting into sci-fi stuff We need severe reality check for current state.

00:01:17: Yeah because the realty kind of stark isn't it?

00:01:20: It really is.

00:01:21: Aaron Prather actually highlighted this really sobering statistic recently, he pointed out that eighty percent of US manufacturing facilities still have zero automation.

00:01:30: Zero?

00:01:31: Like

00:01:32: none?

00:01:32: Zero!

00:01:33: None.

00:01:33: and the hold-up here isn't a lack of like advanced AI models or fancy robotics... Right The bottleneck is pure integration.

00:01:41: I mean a factory has this highly complex deeply entrenched physical system, you can't simply push an over-the-air update to a twenty year old stamping press and expect it suddenly wake up.

00:02:07: Right,

00:02:07: I saw that.

00:02:08: Yeah and they're pushing past just basic analytics into what they call systemic AI.

00:02:13: so there trying to move the industry from advisory systems where like a dashboard passively tells a human operator that a motor is running hot-to true agentic manufacturing

00:02:25: which is a fundamental shift in control if you think about

00:02:29: it how's?

00:02:29: oh well

00:02:30: In an agentics system The AI actually senses the temperature spike decides on the optimal intervention based on the current production schedule and acts completely autonomously.

00:02:43: Oh, wow!

00:02:43: Yeah it literally throttles the motor down or reroutes the workflow itself.

00:02:47: It just completely removes human latency from the decision loop.

00:02:51: But wait if eighty percent of factories have zero automation Isn't pushing agentic AI into these legacy facilities essentially like, I don't know dropping a Ferrari engine in to horse drawn carriage?

00:03:02: Yes.

00:03:03: That's exactly it

00:03:03: because the underlying chassis of factory just wasn't designed.

00:03:06: handle autonomous high-speed decision making right

00:03:09: precisely.

00:03:10: and that structural mismatch is Exactly why so many of these ambitious ai pilots just crash and burn.

00:03:17: Ravi Kunju provided some really great analysis on Samsung's twenty-thirty AI factory vision.

00:03:23: Right,

00:03:23: Samsung is going all

00:03:24: in?

00:03:25: Yeah they're committing to fully autonomous AI across every single factory globally by twenty thirty.

00:03:30: but to attempt that you know organizations usually defaults just dumping their historical factory data into a massive centralized Data Lake assuming the AI will figure it out

00:03:40: which never works

00:03:41: Never!

00:03:42: Kunju points out this specific data lake approach causes seventy percent of pilots fail.

00:03:47: That's huge!

00:03:49: Data lakes work perfectly fine in the IT world, don't they?

00:03:52: Why do they fail so spectacularly in operational

00:03:54: technology?".

00:03:55: Well

00:03:55: because a data lake is fundamentally just massive dumping ground for raw unstructured information.

00:04:00: Think about it like this... If you feed an AI model the number seventy-five from an IT database It might clearly be user count.

00:04:08: But if you pull seventy five from a factory data lake, the AI has absolutely no idea.

00:04:14: If that represents a temperature reading in Fahrenheit of vibration frequency and Hertz or like a specific quantity of defective parts

00:04:22: would lacks operational context completely

00:04:24: exactly.

00:04:24: Kunju emphasizes that You need a semantic layer specifically utilizing knowledge graphs.

00:04:30: So what is a semantic?

00:04:31: Layer?

00:04:32: actually mechanically do to that number seventy-five before they even touches it.

00:04:37: It maps the relationships.

00:04:39: So, The semantic layer takes that raw telemetry data and attaches it to a specific sensor.

00:04:44: Right And that sensor is attached with a specific spindle which part of a specific CNC machine Which currently running in a product

00:04:52: batch I see.

00:04:53: so it structures the data

00:04:55: right?

00:04:55: So the AI agent inherently understands the actual physics and hierarchy on factory floor.

00:05:00: It transforms a totally meaningless number into actionable context.

00:05:04: And what's fascinating is that this level of architecture isn't just walled off for giants like Samsung and Stellantis anymore.

00:05:10: No, not at all!

00:05:10: Benjamin Karash was discussing Corello recent deployments... ...and they are putting private AI assistants in to mid-sized manufacturers now Which is

00:05:18: awesome

00:05:18: It is.

00:05:19: These assistants are pulling from the ERP, the MES live machine data and even the undocumented tribal knowledge of the senior operators.

00:05:27: The

00:05:27: Tribal Knowledge is huge.

00:05:29: Yeah.

00:05:29: so a floor manager at a medium-sized plant can just query their AI assistant about specific bottleneck And it instantly cross references supply chain data with live spindle data to give contextualized answer.

00:05:45: But, and this is a big but attempting to integrate that level of intelligence directly into physical moving factory floor is inherently risky.

00:05:54: Imagine writing a routine software update that accidentally sends a two-ton robotic arm smashing straight through a multi million dollar conveyor belt.

00:06:02: Because in standard software, a bug just means an app crashes right?

00:06:05: Exactly!

00:06:05: In manufacturing... A bug means physical destruction massive downtime and serious safety hazard.

00:06:11: So if the AI makes bad decisions based on faulty semantic map machinery literally breaks

00:06:16: Which brings us to next massive trend.

00:06:19: because of extreme risk profile The industry is moving toward an absolute requirement.

00:06:24: You have to build the system virtually before you pour a single ounce of concrete.

00:06:28: Digital twins and simulations?

00:06:30: Yes,

00:06:31: mechanics testing the Ferrari engine in virtual wind tunnel... Michael Walker and Matthias Heinecke both highlighted the really verifiable success Posicot is experiencing with this right now.

00:06:43: Oh using The Siemens digital twin composer?

00:06:46: Exactly integrated with NVIDIA Omniverse.

00:06:48: Yeah, they aren't just building a nice you know three-D visualization.

00:06:52: It's not a video game Right.

00:06:53: They are constructing high fidelity physics based simulations of their factory lines.

00:06:59: Every single conveyor belt, every robotic palletizer, every fluid dynamic in the bottling process.

00:07:05: it's all simulated to behave exactly as it would under real-world physical

00:07:09: constraints.".

00:07:10: And...the financial impact doing that is staggering!

00:07:12: I mean by running these AI driven simulations at the omniverse before they commit capital into the physical build, PepsiCo has catching ninety percent potential integration issues early.

00:07:22: Ninety percent?

00:07:23: That's

00:07:23: insane!!

00:07:24: Yeah They're cutting their capital expenditure by ten to fifteen percent, and boosting throughput by twenty percent.

00:07:30: They literally know the factory will hit its production targets before they even order the equipment.

00:07:35: It changes entire paradigm of capital deployment.

00:07:38: You are no longer just crossing your fingers during physical commissioning phase

00:07:41: Exactly.

00:07:42: But it's not for macro-level facility planning.

00:07:45: It works down the micro level too Like single part manufacturing.

00:07:49: Oh you mean two action man machine tools example.

00:07:52: Yes

00:07:53: Sasha Fisher shared this brilliant use case.

00:07:56: Yeah, they needed to manufacture this highly complex high performance golf putter.

00:08:00: and

00:08:00: usually Machining something with that level of precision takes so much trial-and-error on a physical CNC machine.

00:08:07: Right you write the g code You run it?

00:08:10: You probably break a tool you scrap the metal And do rewrite the coat

00:08:13: massively inefficient iterative process.

00:08:16: It just eats up raw materials an expensive machine time.

00:08:19: but instead they use a Siemens CNC digital twin workflow.

00:08:24: They took the exact kinematics of their specific machine, the exact tooling parameters and validated the entire machining process in the Digital

00:08:32: World.

00:08:33: They optimized toolpaths completely virtually.

00:08:36: Yep!

00:08:37: And by that time code finally reached the physical machine...they had cut the cycle-time by seventeen percent.

00:08:43: zero physical scrap Zero guesswork on floor.

00:08:46: That is so cool.

00:08:47: But wait, let me ask you this to build a digital twin that accurate whether it's an entire PepsiCo bottling line or just one CNC machine wouldn't You have to spend like two or three years Just instrumenting the factory and gathering every conceivable data point first

00:09:02: right?

00:09:02: That's what everyone assumes.

00:09:03: Yeah

00:09:03: I mean by the time you built the Twin The Factory would be obsolete.

00:09:07: well Brent Roberts brought up a critical insight about this.

00:09:10: That assumption is the exact trap that kills digital twin initiatives, okay?

00:09:14: Engineering teams often paralyze themselves by demanding a perfectly comprehensive data ecosystem before they even begin modeling.

00:09:22: but Roberts argues that you do not need all of the data.

00:09:24: You don't No!

00:09:35: So if your goal is just to simulate a bottleneck at the packaging station, you definitely don't need the ambient temperature data from the loading dock.

00:09:42: Right

00:09:43: or historical maintenance logs of the forklift...you don't care!

00:09:46: You isolate specific PLC signals, exact cycle times and layout geometry for that one specific node.

00:09:54: That makes so much sense.

00:09:55: The engineering challenge isn't about hoarding massive volumes.

00:10:00: It is about ruthlessly defining the minimum viable parameters required to make a trustworthy simulation and then just ensuring the integrity of that specific data feed.

00:10:09: But, this is huge.

00:10:10: but if you need that specific date-of-feed be absolutely trust worthy.

00:10:14: You quickly realize that legacy IT infrastructure for a factory Is entirely insufficient.

00:10:19: Oh

00:10:19: completely!

00:10:20: The old way of wiring PLCs To a localized server won't survive demands Of feeding alive.

00:10:25: digital twin or an agentic AI?

00:10:28: No way.

00:10:29: The underlying nervous system has to be fundamentally re-architected, and we're seeing this hard pivot away from traditional software categories.

00:10:37: Michael Finocchiaro discussed the shift.

00:10:40: he noted that manufacturers are no longer just out to buy an MES a manufacturing execution system.

00:10:46: Right

00:10:46: people are moving past the traditional MES.

00:10:48: so what are they doing instead?

00:10:49: They're having MT-stack.

00:10:53: Mundio.

00:10:54: Yeah,

00:10:54: MNE standing for Manufacturing Intelligence and Network Transformation.

00:10:58: Okay.

00:10:58: so why is the traditional MES falling short here?

00:11:01: Because a traditional Mes was often just implemented in total silo.

00:11:05: Modern intelligence driven projects fail because there's this constant friction over data ownership.

00:11:10: Like who actually owns The Truth?

00:11:12: Exactly Does PLM product life cycle management system own the definitive bill of materials Or does the ERP own it?

00:11:20: Does the MES dictate the workflow, or is there a new AI agent.

00:11:24: That sounds like nightmare to unravel!

00:11:26: It IS The MINT stack philosophy forces an organization to define clear unified data ownership across all these architectural layers Because without that you basically just get competing databases overriding each other

00:11:42: And the integrity of this data within the stack Is so incredibly fragile.

00:11:47: David Schultz actually shared a warning about this.

00:11:49: that mechanically explains why so many machine learning models end up failing on the factory floor.

00:11:54: Oh,

00:11:54: the telemetry versus event data issue?

00:11:56: Yes

00:11:57: he focused on the critical difference between raw telemetery data and human-event data And why embedding them together is literally catastrophic.

00:12:05: This is probably the most common architectural stake in industrial data engineering right now.

00:12:09: Yeah

00:12:10: think about the mechanics of it.

00:12:11: Telemetry Data is objective factual machine physics.

00:12:14: Right

00:12:14: The motor is drawing this exact amount of current, the spindle is rotating at this exact RPM.

00:12:20: It's just

00:12:20: math!

00:12:21: Exactly event data though...is purely contextual.

00:12:24: it's an operator physically punching a code into a screen saying the line is down because of material jam.

00:12:29: but what happens when maintenance tech realizes three hours later that wasn't a Jam at all actually a faulty optical sensor?

00:12:38: well if your architecture has permanently fused that initial incorrect human event code directly into the time series telemetry data, you've fundamentally corrupted the historical record.

00:12:50: Yes

00:12:51: it's okay.

00:12:52: It's exactly like taking raw security camera footage and permanently editing the video file based on what an eyewitness thinks they saw in the moment.

00:13:00: Oh

00:13:00: that's a great analogy Right

00:13:02: The factual tape is completely ruined.

00:13:04: And when you feed that corrupted tape to machine learning model

00:13:13: Which completely defeats the purpose of AI, you have to store telemetry and event data in separate parallel databases that are linked relationally.

00:13:21: Exactly!

00:13:22: The event context can be debated over time but raw physics must remain totally pristine

00:13:30: Absolutely, and you know beyond data integrity this new nervous system also has to handle just an incomprehensible level of speed.

00:13:36: Oh the speed is wild!

00:13:39: Jacob Abel recently posted the results of a stress test he ran on OT's virtual PLC The VPLC

00:13:45: Right?

00:13:46: To feed these digital twins and AI models, you have to constantly pull the machines for their status using OPC UA tags.

00:13:54: Now a traditional highly tuned physical PLC might handle maybe three thousand to five thousand tags per second

00:14:00: Which used be plenty fast Used

00:14:01: to be.

00:14:02: Abel's test of the virtual PLC was delivering forty-thousand to one hundred thousand Tags Per Second.

00:14:07: Wow!

00:14:08: That is staggering leap in throughput.

00:14:10: It really highlights the physical limitations of legacy hardware.

00:14:13: I mean, when an AI agent is making autonomous sub-second decisions about a high speed manufacturing process three thousand updates per second just creates way too much latency.

00:14:23: So the control logic literally has to be abstracted into a software-defined environment just to achieve the compute scale required.

00:14:30: Okay, so we have AI models and high fidelity digital twins And this hyper fast structurally sound data architecture.

00:14:38: But at the end of day all these digital signals have to translate in physical mass actually moving through a facility.

00:14:44: Right!

00:14:45: The real world.

00:14:46: How are actual robots and human operators standing next to them adapting?

00:14:52: Well, the deployment of robotics is becoming significantly more pragmatic I would say and collaborative.

00:14:59: The industry is really moving away from dangerous robots locked inside isolated safety cages.

00:15:07: Oh, for sure

00:15:08: we're moving toward flexible hybrid production environments now.

00:15:12: Robert Little shared a really prime example of this with Amazon's Vulcan system

00:15:16: one that just won Robot Of The Year.

00:15:17: Yeah

00:15:17: exactly.

00:15:18: and the mechanical design philosophy behind Vulcan is fascinating.

00:15:22: it does not exist to replace a human picker.

00:15:24: It was engineered specifically to handle the ergonomically destructive tasks.

00:15:28: oh

00:15:29: right reaching high and low.

00:15:30: yeah Vulcan extends to reach the highest storage bins and drops down to access the lowest bins while leaving fast, comfortable ergonomic center zone picking entirely into human employees.

00:15:42: That makes perfect sense!

00:15:43: It's a highly optimized division of physical labor based on actual strengths each entity.

00:15:48: Exactly.

00:15:49: But you know... getting A robot like Vulcan, or even a basic collaborative robot integrated into a mid-sized facility brings us right back to the integration bottleneck Aaron Prather warned about at the very beginning of this discussion.

00:16:02: The

00:16:03: integrator squeeze.

00:16:04: Yes Victor Alvarez refers to this barrier as...

00:16:09: So why exactly are small and medium enterprises getting squeezed out here?

00:16:13: because the hardware cost of the robot is only a fraction of total expense.

00:16:17: Right, everyone forgets about the rest.

00:16:19: The real costs are custom engineering, multi-week onsite visits from integration specialists safety validations and honestly the unpredictable downtime during installation.

00:16:29: Those integration costs frequently price SMEs out at the automation market entirely.

00:16:33: But AI we discussed earlier now being pointed directly to that integration problem which is incredible.

00:16:39: Oh yes, Lukas M. Ziegler highlighted how Robotic recently launched an AI platform called IQ that automates the integration engineering itself...

00:16:48: Which is mind-blowing!

00:16:50: It is you no longer need a specialist to fly out for a site visit?

00:16:54: The AI assesses the digital footprint of your floor mathematically validates the cycle times and simulates the entire robotic work cell virtually.

00:17:03: So it basically makes the deployment of a physical robot nearly as predictable and standardized, like provisioning a new server.

00:17:11: Which is massive leap for... democratizing automation.

00:17:15: Yeah, however and there is always a how ever in manufacturing we have to look at the reality of maintenance.

00:17:21: yeah what happens when it breaks?

00:17:22: exactly?

00:17:23: When you tightly couple ERPs MES layers virtual PLC's semantic graphs and physical robotics into one unified cyber-physical system The complexity of troubleshooting skyrockets

00:17:33: right because when the line stops You can't just grab a wrench anymore.

00:17:37: no And you can't Just blame this software either.

00:17:39: It's all connected.

00:17:40: Well, Alana Murray laid out this brilliant framework for surviving this exact complexity.

00:17:45: She detailed a specific mental chain that maintenance and engineering teams must rigorously follow when diagnosing a fault in these environments.

00:17:53: And you have to walk the chain.

00:17:54: an order right?

00:17:55: Exactly it forces illogical progression rather than just guessing.

00:17:58: So step one process.

00:18:01: You look at the raw physics first.

00:18:04: Is a physical valve actually stuck?

00:18:06: is a pipe blocked?

00:18:07: okay check The real world first yes.

00:18:09: Then step two, instrumentation.

00:18:12: Is the physical sensor actually installed correctly and is the physical reading it's outputting even mathematically possible?

00:18:18: Right

00:18:18: then what?

00:18:19: Step three is control.

00:18:21: you look at the PLC logic.

00:18:22: did a loop fail or threshold get crossed incorrectly?

00:18:25: And finally step four network.

00:18:28: are the OPC UA tags dropping packets?

00:18:30: is the data lake failing to ingest?

00:18:32: so You have to trace this symptom linearly from the physical steel all the way up to the cloud

00:18:37: Precisely.

00:18:38: You cannot skip a link in that chain because, In A Fully Connected Factory, Isolated Failures Essentially Don't Exist Anymore.

00:18:45: Yeah Tony LaRoy articulated this beautifully when he discussed the extreme danger of just one small change.

00:18:51: Oh The Ripple Effect

00:18:52: Exactly!

00:18:54: A production manager might request what appears to be a microscopic mechanical tweak To a conveyor guide.

00:18:59: rail Right?

00:19:00: Seems simple enough

00:19:01: Right, but moving that physical rail means the timing of part arriving at sensor changes and time change requires an update to core PLC logic.

00:19:10: And updating it alters safety condition parameters?

00:19:13: Exactly!

00:19:14: Which then require HMI screens to be redesigned so operators know what new fault codes mean which triggers a mandatory revalidation for the entire cell.

00:19:26: So a physical adjustment measured in literally millimeters creates a massive, expensive tsunami across the digital invalidation layers.

00:19:35: The hardware and software are entirely fused And

00:19:38: that fusion is exactly why the transition to smart manufacturing is just brutally difficult.

00:19:43: It requires an organizational respect for the absolute interconnectedness of mechanical physics... ...the digital architecture....and human operators.

00:19:51: Yeah, I mean from seventy percent failure rates due to raw data lakes To the necessity of building virtual PepsiCo plants down to the actual mental discipline required to troubleshoot a single sensor.

00:20:02: The overarching theme here really is that?

00:20:04: The foundation has to be flawless before the intelligence can scale.

00:20:08: I couldn't agree more and to That point i want to leave you the listener with one final thought regarding where this Foundation Is truly being tested right now.

00:20:17: okay

00:20:17: let's hear it.

00:20:18: Nicholas Savage shared an analysis of India's rapidly evolving industrial sector.

00:20:23: You know, in the West The industry tends to obsess over building these pristine highly controlled Greenfield mega facilities from scratch.

00:20:31: right but savages argues that the next true global leader In Industrial AI will likely emerge From a messy chaotic brownfield factory in india.

00:20:41: why is That?

00:20:42: The reasoning Is that A typical legacy environment There might contain Over a Hundred Different completely incompatible brands of ancient PLCs.

00:20:50: They operate under hypercost consciousness and really fragmented supply chains, it is the ultimate unforgiving stress test

00:20:57: I see.

00:20:58: yeah

00:20:58: if an engineering startup can figure out how to capture trustworthy data normalize it through a semantic layer And successfully deploy agentic AI in that specific environment they haven't just solved a localized problem.

00:21:11: They've built the definitive global playbook for legacy integration.

00:21:14: Exactly!

00:21:15: That completely flips the script on how we view legacy factories, they aren't just technical debt... ...they are literally the ultimate crucible for proving out these AI architectures

00:21:25: Spot-on.

00:21:26: If you enjoyed this episode new episodes drop every two weeks.

00:21:29: Also check our other editions of Digital Construction and Digital Power Tools.

00:21:34: Yeah thanks so much.

00:21:34: taking a deep dive with us today.

00:21:36: Absolutely

00:21:37: Make sure to hit subscribe.