A single bearing goes in the gearbox of a finishing stand. The hot strip mill stops. The caster backs up, and the cold mill sits idle for two, maybe three days, waiting for coils. Add it all up and it’s easily a few hundred thousand dollars. Gone. Yet plenty of plants still service their machines “once a month” or simply “when it breaks.” Both habits cost far more than they save.

From Calendar Schedules to Condition-Based Maintenance

Calendar maintenance had its logic. Back when all anyone knew about a machine was its age and the date someone last looked at it, changing the oil every 2,000 hours was a sensible rule. Bearings every 18 months. Done.

The machines never read that schedule, though. Same bearing type, same kind of drive. In one case it runs three times its rated life because the housing is clean and the grease is right. In another it’s finished in four months, because a tired seal let scale and cooling water in. When industrial giants bring in specialized IT services and solutions for manufacturing to link shop-floor sensors with their central ERP and EAM systems, they finally get full control over the real technical state of every shop, from the ore yard to the finishing line. And the vibration alarm? No longer one more email sitting unread till Monday. It lands as a work order, with parts, crew hours and a shutdown slot already in it.

Real plants are messier than textbooks. Most mills juggle several approaches at once:

• Run-to-failure. Fine for a lighting circuit or a backup pump. For a mill main drive? Never.
• Fixed intervals from the OEM manual. Planners love it. Half the parts go to scrap perfectly healthy, and early failures slip through anyway.
• Condition-based. Vibration crosses its alarm line, the oil gets dirty, a winding runs hot. Then the crew goes in, not before.
• Predictive. Trend data plus a model that guesses how many weeks a part has left, so the repair gets parked in the next planned stop.
• Prescriptive, the newest kid on the block. It also says what to do: more grease, a realignment, a lighter load, or “order the bearing today.”

Why bother? Look at the P-F curve. P is the moment a defect first shows up in the data, F is the moment the machine actually fails. With vibration analysis, a bearing gives warning weeks ahead, sometimes months. Temperature buys days. Someone’s ear, a few hours if lucky. More warning, less panic. Makes sense, doesn’t it?

What the Sensors Actually Catch

Vibration Analysis on Mills and Gearboxes

Vibration still does most of the heavy lifting. Accelerometers get bolted straight onto bearing housings. Fans and conveyor drives up on the gantries (the ones nobody wants to climb to in January) get wireless IIoT sensors. Big turbine shafts have had proximity probes for ages. The overall velocity number, checked against ISO 20816 zones, tells the team something is off. The spectrum tells them what.

Here’s how it looks on a plate mill. An analyst pulls up the pinion stand data and spots sidebands creeping up around the gear mesh frequency, spaced at the output shaft speed. Textbook sign of a cracked or eccentric gear. Catch it at this point, and a crack in the shaft or pinion shows up three weeks or more before it would snap. Three weeks is a lot. Enough to get a spare delivered and swap it during the next roll change, with nobody running around at 3 a.m.

Bearings snitch on themselves too. Envelope analysis catches the outer- and inner-race tones, BPFO and BPFI in analyst slang, while the overall number on the screen hasn’t even twitched.

Temperature and Thermography: Late, but Not Useless

For bearings, temperature is late to the party. By the time it climbs, the damage is done. Elsewhere, though, nothing replaces it. RTDs in motor windings and thrust bearings. Infrared cameras on switchgear, busbars and ladle-car electrics. Shell scanners on kilns and BOF converters. A hot patch on a converter shell means the refractory is getting thin, and that’s not something to “keep an eye on.”

Oil and Lubricant Condition

Oil keeps a diary. Every gearbox, every AGC hydraulic unit, every turbine lube circuit scribbles in it, and inline sensors now read along, live:

• how dirty it is (ISO 4406 codes) and whether the dirt is steel;
• water in ppm, since a few hundred ppm in turbine oil already eats into bearing life;
• viscosity drift and oxidation;
• oil temperature, plus pressure drop across the filters.

Months of flat readings, then ferrous particles jump. Nine times out of ten, a gear tooth or bearing race has started to flake. The lab confirms it with ferrography. Put that next to the vibration data, and “something’s wrong with gearbox 3” becomes “inner race, bearing 2, replace at the next stop.”

Ultrasound and Motor Current Signature

Two quieter tools fill the gaps. On slow bearings vibration says almost nothing, but ultrasound can literally hear the friction and tell the fitter with the grease gun when to stop. Sounds trivial? It isn’t. Over-greasing wrecks bearings just as reliably as starving them. Then there’s motor current signature analysis: broken rotor bars, eccentricity, all read off the drive cabinet. No crawling around a hot motor in some dusty corner.

Mills, Turbines, Heavy Conveyors: Where Things Really Break

Copying a monitoring plan from another plant rarely works. Each type of equipment breaks in its own way. In steel and heavy machinery, the usual suspects are:

• Rolling mill stands. Chock bearings on work and backup rolls, spindles, pinion stands, main gearboxes. Every threading and every roll change hits them with shock loads.
• Turbines and blowers: blast furnace blowers, top-gas recovery turbines, steam turbines in the plant’s own power station. Proximity probes to API 670 and Bently Nevada System 1 have been around here for decades. The new part is getting that data out of its silo.
• Heavy conveyors. Ore and sinter belts kilometers long, with thousands of idlers. One seized idler can slice a belt that costs more than the whole monitoring setup.
• Fans and pumps. ID fans in sinter and dedusting plants cake up with dust, lose balance, start to shake. A caster’s cooling water pump? Lose it, and the first call goes to safety, not production.

SCADA, IIoT, EAM: Who Does What in the Analytics Layer

A sensor on its own fixes nothing. Somebody has to act. The usual split looks roughly like this. SCADA and DCS keep the process running. A historian, often AVEVA PI, hoards years of signals. The IIoT platform swallows the fast condition data. And the EAM or CMMS is where all of it finally becomes a job with a name on it.

Plenty of mature tools out there, and a big plant tends to end up with two or three:

• Siemens Insights Hub (MindSphere, before the rename). A natural fit on floors packed with Siemens drives.
• PTC ThingWorx, frequently bolted to Kepware so it can talk to every odd PLC in the building.
• GE Vernova APM, the old GE Digital APM. Big on rotating equipment.
• IBM Maximo Application Suite. Monitor, Health, Predict, one roof.
• SAP EAM or S/4HANA Asset Management, if maintenance has to sit in the same ERP as purchasing.
• Honeywell Forge, for asset health on the operations side.

And somewhere between all of these sit integrators like DXC Technology. Their job is the plumbing: OT talking to IT, data pipelines, IEC 62443 security, and an anomaly score from a vibration model popping up as a high-priority notification in SAP or Maximo. That last bit? Most pilots die right there. Quietly. Lots of plants have nice dashboards. Very few have dashboards that reserve the spare part on their own.

Metrics That Prove the Strategy Works

Nobody in the boardroom wants to look at a spectrum. They want numbers. Five of them, mostly:

• OEE, Overall Equipment Effectiveness: availability, times performance, times quality. 85% is the usual “world-class” mark, while plenty of heavy-industry lines are stuck somewhere in the 60s.
• MTBF goes up, MTTR goes down. Why shorter repairs? Boring answer. The crew walks in with the parts already on the cart.
• Share of planned work. Mature reliability teams aim for 80–90%.
• PM compliance. Under 90%? Then the plan lives mostly on paper.
• Maintenance cost per ton. The one the board actually reads.

And one number worth keeping in the back pocket. Siemens, in its 2024 “True Cost of Downtime” study, put the yearly bill for the world’s 500 largest companies at roughly $1.4 trillion. Eleven percent of revenue, give or take. Is heavy industry the exception? Hardly.

How to Roll It Out Without Stalling

The usual mistake is going too big. Two thousand sensors, one shiny cloud platform, six months. Then alarms ringing day and night, and operators who’ve stopped looking. Familiar story? A calmer path goes like this:

1. Criticality analysis first (FMECA). Twenty or thirty machines that actually halt production? Start there.
2. Pick the method for the failure mode. The vendor brochure can wait.
3. Record a baseline. Several weeks of normal running before any alarm goes live.
4. Connect alerts to the EAM workflow on day one, not “in phase two.”
5. Train the reliability engineers. ISO 18436 for vibration analysts isn’t a bad yardstick.
6. Once a month, sit down with the false alarms and move the thresholds.

Scaling across the whole shop? After the first machines behave. Not earlier.

Final Word

So. Unplanned downtime in a steel plant is hardly ever sudden. The bearing was getting warm. The oil was darkening. The spectrum had been growing sidebands for weeks. Someone just needed to be listening. Plants that join sensors, analytics and maintenance planning into one loop stop putting out fires and start planning their stops. Isn’t that what maintenance was supposed to be in the first place?