What to Fix First When a Local Factory's Efficiency Depends on Your Applied Physics

You walk into the factory floor. The plant manager says, 'We're losing 15% throughput. Your job is to find it.' You have a multimeter, a thermal camera, and maybe an oscilloscope. Where do you point them first? This isn't a textbook problem—it's a mess of vibration, heat, and noise. Here's the thing: most new applied physicists start with the most visible problem, like a motor that's clearly overheating. But that's often a symptom. The real fix is buried in a sensor that's been drifting for months, or a heat exchanger whose fins are caked with dust. Let's walk through the logic of triage, with the specific trade-offs you'll face.

The First Hour: Triage on the Floor

Reading the Room—Operator Interviews

Your boots hit the concrete. The line is running, but something's off—maybe a vibration you can feel through the soles, maybe the way a conveyor hesitates at the transfer point. Before you touch a single instrument, find the operator who's been watching this line for three years. Ask one question: “What changed last week?” I have seen entire thermal hunts collapse because nobody asked the night crew about the new lubricant they swapped in. Operators know the machine's personality—the clatter that means a bearing is loosening, the sweet spot where a sensor reads true. Most engineers skip this step; they jump straight to data sheets. That's a mistake. The first hour isn't about precision—it's about signal. You're looking for the one person who says, “Funny you mention that…”

Worth flagging—operators rarely speak in Feynman diagrams. They speak in complaints. “The seam keeps blowing out around 2 PM.” That's your clue. It points to a thermal drift or a humidity shift, not a quantum effect. So listen for the pattern behind the complaint. Then verify with your own eyes. Stand at the machine for three minutes, no clipboard, just watching. The catch is that most physicists hate doing this—we want meters, not stories. But the meter lies if you point it at the wrong node. The story tells you where to point.

Quick Thermal Scan Protocol

Now you have a suspect zone. Grab the thermal camera—but don't shoot the whole factory floor. That wastes battery and buries the signal. Instead, run a 10-second scan of the transfer point, the motor housing, and the bearing flange at the drive end. Three targets. Why those? Because friction generates heat before it generates failure. A 5°C gradient on a bearing surface? That's your signature. A uniform temperature rise across the whole machine? That's the ambient air, not the problem. False leads kill triage. I have wasted forty minutes on a hot cable tray that turned out to be a steam line underneath. The thermal scan is a sieve, not a diagnosis. It tells you where not to look.

One more thing: log the ambient temperature and humidity before you start. Not because you'll use it now—you won't—but because when you check the data four hours later, you'll need to separate machine drift from weather drift. That's a cheap insurance policy. Most teams skip it. Then they chase ghosts all afternoon.

Logging Baseline Power

Plug a power logger into the main drive motor. Not the control cabinet—the motor itself. Why? Because the control cabinet shows you setpoint; the motor shows you reality. A 15% dip in current with no change in speed means the load dropped. Something upstream is starving the machine. Or a belt is slipping. Or the material changed. Baseline power is your ground truth, and it takes two minutes to set up. Let it run while you do everything else. When you come back, you have a time series, not a guess.

But here's the trap: power alone doesn't tell you efficiency. A motor pulling 30 amps at rated speed might be fine, or it might be fighting a seized bearing. Current is a symptom, not a cause. So pair the power trace with your thermal snapshot and the operator's story. Triangulate. That first hour is not about solving the problem—it's about building a shortlist of three suspects. The physics comes later. Right now you're a detective with a thermistor and a healthy dose of skepticism. That's enough.

One rhetorical question worth asking yourself before you leave the floor: Am I measuring what I think I'm measuring, or am I measuring what's easy? The honest answer usually points you toward the second sensor you forgot to bring.

“The machine told me nothing until I learned to ask the wrong questions first. Then it lied in a way that made sense.”

— maintenance lead, mid-sized extrusion plant, after a three-hour power chase

Foundations Most Physicists Skip

Sensor calibration drift

The first assumption I almost always see on a factory floor is that the sensor is telling the truth. A temperature probe reads 72°C, the line supervisor shrugs, and everyone adjusts the process around that number. But that 72°C might be 68°C—or 78°C—after six months of thermal cycling, dust accumulation, and the occasional forklift bump. I once watched a team chase a 5% yield drop for three weeks. They replaced pumps, recalibrated flow controllers, even rewrote the PLC logic. The fix? A $40 thermocouple that had drifted 11°C off nominal. The catch is that most preventative maintenance schedules treat sensors as permanent fixtures, not consumable components. They drift. They age. And unless you have a field calibration log with actual dates—not just a sticker that says 'Calibrated 2022'—you're debugging ghosts.

Worth flagging: many applied physicists skip the simple check. They jump straight to modeling the heat transfer or the fluid dynamics. But a model built on bad input is just expensive fiction. So before you open any simulation software, put a handheld meter on that probe. Compare three points across the operating range. The error will surprise you—not because the sensor was cheap, but because drift is silent.

'The machine says 100 kN. The part breaks at 95 kN every time. The machine is the one lying.'

— plant engineer, overheard during a root-cause meeting that ran two hours too long

Ground loops and noise

Most teams skip this: electrical noise can masquerade as a physics problem. A pressure sensor shows wild 60 Hz ripple—teams blame pump cavitation, pipe resonance, maybe a failing impeller. They install dampeners, change pump speeds, add accumulators. None of it works. Meanwhile, the actual culprit is a ground loop between the sensor shield and the VFD drive cable running in the same tray. I have fixed exactly this at three different factories. The fix was a $5 isolation module and rerouting two wires. No physics theory required—just a scope and the patience to look at the signal before the math.

Honestly — most physics posts skip this.

The pattern repeats with thermocouple leads near motor windings, with strain gauges next to welding equipment. The human brain wants a physical story—vibration, heat, pressure—because those feel solvable with equations. But the electrical story is often cheaper and faster to rule out. A quick litmus test: disconnect the sensor, terminate with a known resistor, and watch the reading. If the noise vanishes, the problem is in the wiring, not the process. That's a five-minute diagnosis that saves three days of CFD simulations.

Nameplate ratings vs. actual load

Nameplate data is a polite suggestion, not a specification. A motor rated for 10 kW at 1800 RPM will deliver that—under lab conditions, at sea level, with perfectly balanced voltage, and after exactly zero hours of run time. Real factories have voltage sags, harmonic distortion, ambient temperatures of 45°C, and belts that are slightly too tight. The actual load curve is different. I have seen a '15-ton' chiller that could only move 11 tons because the condenser coils were partially fouled and the refrigerant charge had leaked down. The maintenance log said 'checked annually.' The chiller had not been topped off in three years.

The pitfall is that engineers trust the sticker because it's printed, official, bolted to the machine. But nameplate ratings are design targets under ideal conditions. They're not performance guarantees. The smart move is to measure actual current draw, actual torque, actual flow rate—right now, with the machine running its typical cycle. Compare those numbers to the sticker. The gap tells you what margin you actually have. And if you're building a physics-based efficiency model on nameplate data alone, you're modeling a machine that never existed.

That hurts. But it's fixable. One afternoon with a clamp meter and a data logger will give you real numbers. Use those. Ignore the plaque. Then your foundation holds.

Patterns That Actually Hold Up

80/20 Rule on Driveshaft Losses

Start with the rotating parts. I have walked into factories where teams spent weeks chasing mysterious efficiency dips—only to find the culprit was a single misaligned driveshaft coupling. The rule of thumb: 80% of mechanical transmission losses come from 20% of joints, usually the ones nearest the motor or the heaviest load. You measure surface temperature at each bearing housing with a contact probe—anything above ambient plus 15°C demands immediate shimming. That sounds fine until you realize most technicians skip the third joint from the motor because it's buried behind a guard. That is where the loss hides.

The catch is that replacing a coupler introduces its own alignment error. We fixed this by marking the shaft flange runout before unbolting—gives you a baseline tolerance of 0.05 mm. Anything worse and you're bleeding 3–5% of input power as heat. Not catastrophic per shift, but over a 16-hour run that's one lost batch every three days. Wrong order. Most teams pull the coupling first, then measure. Instead: measure while loaded, mark orientation, then pull. Saves an entire afternoon of rediagnosis.

Thermal Gradient Mapping

Thermography gets misused daily. Operators point a camera at the hottest spot, flag it, and move on. The real signal is the gradient—the difference between adjacent zones across a process line. A 12°C drop over 30 cm on an oven wall means the insulation has delaminated; the same drop on a chilled water pipe means a flow restriction. I once traced a 7% yield loss to a 4°C variance in a curing tunnel that three previous audits had called "within spec." They were reading absolute values, not gradients.

Worth flagging—gradient mapping works only if you log the ambient baseline first. A humid August afternoon can shift absolute readings by 6°C and fool you into rebalancing HVAC when the real fault is a stuck damper. Do this: walk the line at midnight when ambient is stable, snap a thermal panorama, then subtract that baseline from your daytime readings. The delta is your actual efficiency leak. That hurts, because it takes an extra two hours nobody budgets for.

'We don't have time to map the whole line.' — said every plant manager before the third overtime shift.

— overheard at a Midwest packaging plant, where gradient mapping later recovered 11% line speed.

Motor Current Signature Analysis

Current draw tells you what torque is doing, not what the motor feels like. Most vibration analysis catches bearing faults after they worsen; current signature catches rotor bar cracks and eccentric air gaps before the bearing sees the imbalance. The pattern: a 1.5% sideband around line frequency (50 or 60 Hz) indicates a broken rotor bar. That bar will fail completely within 200 operating hours. Not yet. But if you catch it at a 1.2% sideband, you have two weeks to schedule the swap—not an emergency shutdown.

The trade-off: MCSA gear is cheap (under $300 for a clamp-on meter with FFT export), but the interpretation requires practice. One production supervisor I worked with flagged every 2% and 3% harmonic as a fault—those are normal slip harmonics. The actual signature is a pair of sidebands spaced at 2× slip frequency. Miss that, and you replace a perfectly good motor while the real bad one spins another month. That's the anti-pattern the next section covers, but for now: learn to spot the 2× slip pair. It's the difference between a planned swap and a burnt winding at 3 AM.

Anti-Patterns and Why Teams Revert

Adding filters before checking grounding

You see the noise on the sensor feed and your first instinct is to slap a low-pass filter on it. I have done this myself — twice on the same machine, actually. The catch is that filters hide the symptom while the root cause sits on the floor, literally. A loose ground strap on a VFD drive can inject 60 Hz hum that looks exactly like process variation. You filter it, the line chart smooths out, and everybody claps. Three weeks later the bearing fails because that unfiltered electrical noise was cooking the encoder anyway. The real fix took twenty minutes: tighten the ground lug, reroute the signal cable away from the motor leads. But teams revert because filtering is a software change — no work order, no maintenance shutdown, no arguing with the plant manager. It feels like progress. It isn't.

Tuning PID loops without fixing mechanical backlash

Worth flagging — this one is painfully common. A conveyor section overshoots its position by 12 mm. The controls engineer cranks up the proportional gain and adds a touch of derivative to dampen the oscillation. The plot looks tight in simulation. On the floor, the motor now hums at a higher pitch and the belt still lands 4 mm off. Why? Because the gearbox has 3 mm of backlash and the coupling is worn. No amount of loop tuning can compensate for metal that moves before it engages. You chase the integral term, you add feed-forward, you try gain scheduling — and each adjustment makes the motor work harder without fixing the slop. What usually breaks first is the motor driver: it overheats and trips. The team reverts to the old sloppy tune because that at least kept the machine running. I have seen this exact pattern on three different factory floors. Fix the mechanical backlash first — then the PID gains work like the textbook promised.

Over-instrumenting without a hypothesis

Slap a thermocouple here, a vibration sensor there, a laser micrometer on the exit. Now you have data. Great — data about what? The common pitfall is believing that more measurements automatically yield more understanding. Wrong order. Without a specific question — 'is the cooling rate causing the seam to blow out?' — you drown in spreadsheets and nobody looks at the numbers after week two. The team eventually disconnects the extra sensors because they trigger false alarms or generate noise that masks real problems. That hurts: the plant spent $12,000 on instrumentation that now sits dark.

Odd bit about physics: the dull step fails first.

'The best sensor is the one that answers a question you wrote down before you ordered it.'

— maintenance lead at a food-packaging plant, after ripping out fourteen unlabeled temperature probes

Instead, pick one variable, form a clear hypothesis — 'if we reduce quench temperature by 5°C, the weld strength increases' — then instrument to test that single link. You can always add more later. Most teams revert to relying on the operator's thumb-and-feel method because those twelve sensors never told them something actionable. Start small, ask a sharp question, measure only what answers it.

Maintenance, Drift, and Long-Term Costs

Compounding Efficiency Loss Per Year

A 2% efficiency dip today? Most plant managers shrug. They shouldn't. That 2% compounds like unpaid debt. I have watched a factory let a 4% belt-drive misalignment slide for eight months. By month nine the motor bearings were chewing themselves apart—unplanned downtime jumped to 14 hours per quarter. The physics here is brutal: friction scales nonlinearly with misalignment angle. You lose 1.2% in year one, 3.7% in year two, then 7%+ when bearings start shedding metal. The catch is—accountants only flag the electricity bill, not the hidden thermal damage rates. We fixed this by bolting an accelerometer onto the drive housing. Cost: $220. Payback: eleven days. That sounds trivial, but the drift was costing $1,600 per month in wasted kWh alone.

Thermal Cycling Fatigue

Every start-up and shut-down cycle is a tiny hammer blow to your plant's metal. Think of thermal expansion as a slow-motion earthquake inside the steel. A conveyor frame that heats to 85°C and cools to 15°C daily—that's a 70 K delta—develops microcracks in about 2,200 cycles. One factory we consulted had a 1997-era oven line. They never touched the anchor bolts. After 28,000 thermal cycles the frame had shifted 4.2 mm laterally. That killed alignment on three downstream stations, creating a 5% QC reject rate. The fix? A $90 box of expansion-compatible shims and a quarterly re-torque checklist. Most teams skip this: thermal fatigue is invisible until the seam blows out. Then you lose a shift.

Worth flagging—the differential expansion between dissimilar metals is worse. Aluminum housings bolted to steel frames. Stainless shafts in carbon-steel bearings. The coefficient mismatch drives a slow creep that standard maintenance schedules miss by a factor of three. I once measured a 0.8 mm gap opening between a motor flange and its pump coupling over eighteen months. The plant had scheduled a bearing replacement but never checked the bolt torque. That gap cost them 23 000 litres of coolant over the following quarter. Small numbers. Big mess.

'The cost of drift is never the drift itself. It's the fire you chase because you ignored the smoke.'

— senior reliability engineer, automotive parts plant, 2023 site visit

Sensor Drift Over Time

Type-K thermocouples drift 1–2°C per 1 000 hours at continuous use above 500°C. That's not speculation—that's published material science. After 8 000 hours your PID loop is chasing a ghost. The controller sees 485°C; the actual flame is at 497°C. That 12°C error changes polymer viscosity by enough to degrade tensile strength by 3–5%. Not catastrophic for one batch. Catastrophic for a million-dollar extrusion contract that fails certification. The cheap fix is a two-point calibration every 400 hours. The expensive fix is a re-spooled production line and a pissed-off client. Which one sounds like your plant's PM schedule?

Airtight sensors drift slower. Pressure transducers with ceramic diaphragms hold ±0.5% for about three years. But humidity sensors? Those bastards wander 2–3% RH per year in industrial air. That matters when your drying oven's dew-point control is the difference between a saleable part and a brittle reject. We installed a secondary reference sensor on one line—cost $150—and discovered the primary unit was reporting 18% RH when the reality was 31%. That line had been over-drying for fourteen months. Energy waste alone: $1,800. Scrap costs: another $4,200. Not yet. That's 6% of the line's annual margin—lost to a component nobody touched. Maintenance, drift, and long-term costs form a triangle where the longest side is always the one you stopped measuring. Measure it. Calibrate it. Or prepare for the bill.

When Not to Use Physics-Based Fixes

Process at thermodynamic limit

Some factories run so close to the edge that physics-based diagnostics become useless. I once worked with a foundry whose furnace operated within 3% of the Carnot efficiency for its fuel mix. No tweak I could measure — heat flux, recirculation rate, refractory thickness — would squeeze out more than a fractional gain. The real constraint was chemistry: the ore grade had dropped, and no applied physics fix could refine dirt into concentrate.

You hit this wall when every process variable already sits at its fundamental bound. Temperature? At the material's decomposition limit. Pressure? One bar above vessel rating. The catch is — the team still blames physics for the shortfall. They run FEA on a heat exchanger that's already perfect. They model airflow when the bottleneck is the supplier's batch consistency. Not every inefficiency is a physics problem. Some are market problems. Some are procurement problems. Calling a thermodynamic dead-end a 'physics fix' wastes time and credibility.

How do you recognize the limit early? Check the dimensionless numbers. If the Nusselt number hasn't budged in six months of tweaks, the heat transfer regime is tapped. If the process Reynolds number is already turbulent and pressure drop is at pipe code max, you're done. That sounds technical — but the practical signal is simpler: when the same measurement gives the same number no matter what you adjust, physics has no more to give. Walk away. Go fix the supply chain instead.

Client unwilling to shut down for measurements

Physics-based fixes demand data. Real data — not trend lines pulled from a fifteen-year-old PLC log. But plant managers hate stopping a line. I get it: downtime costs $4,000 per hour, and my thermocouple grid needs the machine cold for eight hours. You can't fit a sensor on a running conveyor without risking a jam. You can't align a laser triangulator on a press that cycles every three seconds.

Worth flagging — I've seen teams burn entire week-long contracts trying to bootstrap measurements from noisy live data. They fit curves to vibration spectra gathered with a smartphone held near a bearing housing. They back-calculate flow from pressure drops across valves that have been throttled manually for years. That isn't physics. That's guesswork with equations. The result is always the same: the model predicts a 5% gain, the client implements the fix, and production drops because the real boundary condition was a worn seal they didn't inspect.

So what do you do? You negotiate a measurement window upfront — or you decline the physics approach entirely. One plant in Ohio let me instrument their dryer during a scheduled holiday shutdown. We collected baseline data in six hours, and the permanent fix held for two years. Another client refused every shutdown request. I handed them a list of low-cost mechanical swaps — replace belts, align pulleys, clean filters — and walked. Sometimes the right call is to do nothing physics-based. Not because it wouldn't work, but because the client won't let you prove it.

Field note: physics plans crack at handoff.

Political or budget constraints

Physics doesn't care about org charts. An electric arc furnace might need a new transformer tap changer — a straight physics fix for power factor correction. But the maintenance budget was spent on the plant manager's pet project last quarter. The CFO sees a $90,000 capital request and kills it, because the board cares about quarterly earnings, not joule losses. You can't model your way past a budget freeze.

I've sat in meetings where the engineering team knew exactly which three variables to fix — duct geometry, refrigerant charge, control loop gain — but the division head had promised the VP that 'no major changes' would happen before the audit. The team asked me to write a report recommending a different filter brand instead. That way they could claim they followed physics advice without actually touching the process. Maddening? Yes. But it happens more often than bad readings.

The pitfall here is over-investing in a physics model that will never get implemented. I now ask three questions before taking any diagnostics contract: Who owns the budget? Will they approve a line stop? And is there a political reason the real problem can't be named? If the answer to any one is 'no', I offer a cheaper, non-physics deliverable — a one-page list of operational quick wins, no sensors required. Better to be honest about constraints than to write a beautiful analysis that gathers dust in a drawer.

“The most accurate model in the world is worthless if the plant floor won't touch it. Physics is a tool, not a mandate.”

— veteran process engineer, overheard after a failed lean-consulting engagement

That quote stuck because it names the real trade-off. You can be right and useless. Or you can be approximately right and actually fix something. When the constraints are political or financial, the applied physicist's job shifts: stop measuring, start negotiating. Or walk away clean — there's no shame in declining a job where physics won't survive contact with the conference room.

Open Questions / FAQ

Can I trust the nameplate?

Almost never. I have walked into factories where the motor rating plate claimed 15 kW but the actual draw sat at 11.3 kW under load—and the line engineer had been chasing a “15 kW problem” for two weeks. Nameplates describe design intent under laboratory conditions, not the drift that happens after three years of dust, voltage sags, and bearing wear. Trust the plate for dimensions and wiring color. For power, torque, or flow rate? Measure it yourself. The catch is that a quick clamp-meter reading at full load might still lie if the supply voltage is low. Always check voltage simultaneously, then compute apparent power. That mismatch alone has killed more efficiency projects than anything else I have seen.

What if the floor is uneven?

Then your vibration data is garbage. Uneven floors introduce low-frequency rocking that mimics imbalance, misalignment, or bearing defects—sometimes all three. Most teams skip this: they bolt a sensor down, log a baseline, and chase a ghost for a month. The fix is cheap. A machinist’s level across the base plate, shims under the feet, re-torque the bolts to spec. Takes forty minutes. Saves two weeks of false positives. Worth flagging—even a 0.5 mm gap under one corner will shift the natural frequency of a 200 kg machine enough that your FFT peaks misalign by 3–8 Hz. That hurts when you're tuning a PID loop to a resonance you don't actually have.

How often should I recalibrate sensors?

Depends on the sensor type and the penalty for being wrong. For thermocouples in a plastic extrusion line? Every six months, because a 2 °C drift changes viscosity and the wall thickness goes out of spec silently. For pressure transducers on a hydraulic press? Every twelve months—unless you see the cycle time creep, then check immediately.

• Accelerometers: every 2000 operating hours or after any crash event.
• Load cells: quarterly; they creep with ambient temperature shifts.
• Flow meters: annual zero-point verification, plus a bucket test if the fluid is dirty.

That sounds expensive. But I have watched a factory spend $12,000 on motor replacements because a $90 pressure sensor had drifted 4 % over two years. The return on a calibration schedule is not sexy—it's boring cash in the bank.

“We calibrated everything once a year. Then we lost a shift to a phantom over-temp fault. The thermocouple was reading 12 °C high.”

— plant engineer, food-processing line, after switching to a six-month cycle

One rhetorical question worth sitting with: can your team spot a 2 % drift in a sensor reading, or will they assume the process changed and start tuning the wrong loop? That's the real failure mode. Most teams revert to three-year calibration intervals because “nothing broke last time.” Then something breaks, and the root cause is a sensor that lied quietly for eighteen months. Maintenance drift is slow. The cost accumulates fast. So pick a schedule, put it on a board, and treat missed dates like missed safety checks—because functionally, that's what they're.

Summary: Three Next Experiments

Log motor current at 1 kHz

Your factory floor probably logs motor current at 1 Hz—once per second, baked into the SCADA trend. That's fine for billing. Useless for physics. A pump that cavitates, a conveyor that skips, a spindle that hunts—those faults live in the 10–100 Hz range. I sat next to an extruder once; the line operator swore the torque felt wrong. The plant engineer showed me the 1-Hz trace. Smooth as glass. We slapped a $40 current clamp on a handheld logger and ran at 1 kHz. Within ten minutes we saw a 12 Hz oscillation in the phase current—vanished in the 1-Hz sample because Nyquist ate it. Set up one channel, log for one production batch, and look for sidebands around the line frequency. You aren't hunting harmonics—you're hunting the missing harmonics that reveal imbalance. The catch: the data file fills fast. A 1 kHz log on three phases eats 300 MB per hour. Pull a 15-minute window, not an eight-hour shift. That hurts—but false negatives hurt more.

Plot temperature gradient across heat exchanger

Most teams look at the outlet temperature and call it done. Wrong order. A heat exchanger's health lives in the difference between hot-side inlet and cold-side outlet—the approach temperature. Plot that gradient over a shift. Does it drift upward? That's fouling. Does it oscillate? That's a control valve hunting or a bypass leaking. We fixed a pasteurizer once where the operators had been cranking steam pressure for months to keep outlet temp at setpoint. The gradient had climbed 12°C. Physics said: fouled plates. The maintenance crew didn't believe it—they flushed it anyway. The gradient dropped back to baseline in two hours. Steam use dropped 18%. Pick one heat exchanger, log inlet and outlet temps on both sides for four hours, and plot the difference. A rising slope means your next fix is chemical cleaning, not a bigger burner.

Ask operator what they've noticed

This one feels soft. It isn't. I have seen physicists walk onto a floor, open a laptop, and ignore the person who has touched that machine every day for four years. That person knows the machine is wrong—they just can't prove it with numbers. They'll say: "It sounds different on Tuesday mornings," or "The seam blows out after lunch." That's a constraint, not a complaint. Most teams revert to pure data because data feels safe. But data without context is noise. The operator sees drift that your 1 Hz log misses. The pitfall: you can't ask "do you see anything unusual?"—everyone says yes. Ask specifically: "When did you last change the settings?" and "What happens right before a reject?" Spend fifteen minutes, notebook only, no laptop. You aren't collecting data. You're calibrating your question.

'The instrument read normal, but the machine sounded like a washing machine full of bolts. I told the engineers for three weeks. Nobody listened.'

— Line operator, food processing plant, 2023