Latest / The Manufacturing Podcast with Fexingo: Factories, Supply Chains, and Industrial Business / How Factories Are Using AI To Stop Waste Before It Starts
Transcript
- Lucas: If these conversations have moved your work forward in some small way, we want to be upfront about how this show stays independent. We deliberately don't run ads on these episodes because we believe the signal to noise ratio matters more than ad revenue. If you want to support that choice, the link is buy me a coffee dot com slash fexingo. Luna: Exactly, and it lets us dig into the messy details without worrying about what a sponsor might think. Thanks for keeping it that way. Lucas: So, let's get into today's angle. For decades, the standard approach to manufacturing quality was reactive. You build the part, you measure it, and if it's out of spec, you scrap it or rework it. The cost of that waste has always been the hidden tax on every factory floor. Luna: Right, and with raw material costs staying volatile through the third quarter of twenty twenty six, that tax is getting harder to absorb. Are we finally seeing a shift toward stopping the defect before it happens? Lucas: We are, but not in the way most people think. It's not just about slapping a camera on an assembly line and using computer vision to spot scratches. That's still valuable, but it's limited by lighting conditions and physical access. The real breakthrough is happening inside the process parameters themselves, using machine learning models trained on sensor data from the machines making the parts. Luna: So instead of looking at the finished product, you're looking at the heat, pressure, and speed data during production? Lucas: Precisely. Let's look at a concrete example from one of the major automotive suppliers we've been tracking. They run thousands of injection molding presses that create complex interior components. Historically, they'd catch dimensional drift only after the part came off the line, which meant wasting plastic, energy, and time for every single bad unit. They deployed an edge computing solution that ingests real-time data from temperature sensors, hydraulic pressure gauges, and cycle timers. Luna: And the model learns what 'normal' looks like for each specific mold, right? So it can flag anomalies before the cycle even finishes. Lucas: That's exactly it. The system identifies subtle patterns that human operators or traditional rule-based alarms miss. For instance, a slight dip in hydraulic pressure combined with a minor temperature fluctuation might predict a short shot—a part that doesn't fully fill the mold—before it occurs. The system then auto-adjusts the injection speed or holds the pressure slightly longer to compensate. This isn't just preventing scrap; it's stabilizing the entire batch. Luna: I've seen similar setups in packaging lines where viscosity changes cause seal failures. The adjustment needs to be millisecond-fast, which is why cloud processing wouldn't work here. Lucas: You hit on a critical constraint. Latency is the enemy. If you send that terabyte of sensor data up to the cloud, wait for inference, and send the command back down, the cycle is already over. This requires edge AI, running lightweight models directly on the PLC or a nearby gateway device. The latency drops to milliseconds, which makes real-time correction possible. Luna: But doesn't training those models require massive amounts of historical data that many older factories just don't have? Lucas: It does, and that's the second hurdle. Most legacy systems log data in silos. The press data is on one server, the quality inspection results are in a spreadsheet on a different machine, and the maintenance logs are in a separate CMMS. You can't train a good model if you don't know which sensor reading corresponds to which successful or failed part. These companies are spending months just cleaning and aligning their datasets before they can even start modeling. Luna: So the bottleneck isn't really the algorithm anymore; it's the data hygiene. That sounds like a nightmare for IT teams in industrial settings. Lucas: It is. And it's also a cultural one. Operators who have been tweaking dials based on feel for thirty years don't necessarily trust a black box telling them to adjust the pressure. We saw a case where a plant manager had to create a feedback loop where the AI suggested adjustments, but the operator had to approve them initially. Over time, as the accuracy proved itself, the trust built up, and the system started acting autonomously. Luna: That makes sense. People need to see the value before they hand over control. Did this particular supplier see significant financial impact yet? Lucas: They reported a twelve percent reduction in scrap rates within the first six months of full deployment. In high-volume injection molding, that's enormous margin improvement. But beyond the direct savings, there's a secondary benefit: consistency. When you stop reacting to defects, you stop chasing fire drills. Your throughput becomes predictable, which helps with scheduling and reduces the need for excess safety stock. Luna: Predictability is just as valuable as yield in today's supply chain environment. If I know my output variance is lower, I don't need to buffer as much inventory. Lucas: Exactly. And this ties back to the broader trend of factories becoming chemical factories or power producers—we discussed that earlier. Now they're becoming data factories. The asset isn't just the machine; it's the intelligence layer sitting on top of it. Investors are starting to price in that digital maturity. Companies that can demonstrate lower operational variance are commanding higher multiples because their cash flows are more stable. Luna: Stable cash flows matter more than ever when interest rates are holding steady at levels that make capital expenditure expensive. Efficiency gains become the primary driver of ROI. Lucas: Right. And we're seeing this expand beyond just manufacturing into logistics and warehousing too. The same principle applies: use sensor data to predict bottlenecks or errors before they happen. But manufacturing is the proving ground because the physics are well-understood. It's easier to model heat and pressure than it is to model customer sentiment or macroeconomic shifts. Luna: True. The variables are controlled. Though I suppose weather affects some raw material inputs even there. But generally, yes, it's a closed system compared to marketing. Lucas: Completely closed. Which is why the lessons from here will likely spread to other industries first. If you can master predictive quality in a steel mill or a chip fab, you can probably adapt it to pharmaceutical production or food processing, where waste is equally costly and regulation is tight. Luna: Pharma is huge on this actually. They have strict compliance requirements, so any automated adjustment has to be fully auditable. Does the AI handle the audit trail? Lucas: That's a key feature. The system logs every suggestion, every operator override, and every automatic adjustment. It creates a immutable record of why a process parameter changed at any given second. For regulated industries, that traceability is non-negotiable. It turns the AI from a mysterious agent into a documented assistant. Luna: Documentation is often the hardest part of new tech adoption. If the tool writes its own report card, it's much easier to sell internally. Lucas: Spot on. And I think we'll see more hybrid models emerging. Not fully autonomous, but augmented. The worker gets a heads-up on their tablet: 'Pressure dropping in Station Four, adjust by two percent.' They verify it visually, and confirm it. It keeps human judgment in the loop while leveraging the machine's pattern recognition. Luna: Augmented reality glasses could play into this nicely. Overlaying the recommended adjustment directly on the machine interface would reduce cognitive load for the operator. Lucas: That's definitely the next iteration. We're moving away from dashboards full of numbers toward contextual, actionable guidance. It's less about giving managers more data and more about giving workers better instructions. Luna: Which brings us back to the workforce aspect we touched on in previous episodes. This isn't about replacing people; it's about upgrading their toolkit. Lucas: Exactly. The skilled technician who understands both the mechanics and the data is going to be the most valuable person on the floor. It raises the bar for hiring, sure, but it also makes the job more interesting and less repetitive. You're solving puzzles rather than just repeating motions. Luna: And if the company invests in that training, retention improves. It's a virtuous cycle. Better tools lead to better outcomes, which leads to better morale. Lucas: I think that's the takeaway for listeners in operations or investing. Look for companies that are treating data infrastructure as core capital, not just an IT expense. The ones who do that well will have lower costs and higher margins as we move through the rest of the year. Luna: Smart move. Keep an eye on the edge computing plays and the industrial software providers who are bridging the gap between old hardware and new analytics. Lucas: Agreed. It's a quiet revolution happening on the factory floor, but the implications for profitability are loud and clear. Thanks for tuning in.