Industrial AI · Maintenance

“The machine always breaks on a Friday night.”

Note on the shift board

AI predictive maintenance

It doesn't break out of nowhere. It has been making noise since Wednesday. We put sensors on it to listen and a model to understand, so the stop happens on Thursday at 06:00, with the part already on the shelf.

Book a Plant X-Ray Watch the week in the Glass Factory

Close-up of an electric motor with a vibration sensor screwed onto the bearing housing and oil around the shaft.

Wed 14:20 · deviation

Among Greek SMEs, 69% say they need predictive maintenance and only 3 to 4% use it. In basic and fabricated metals, AI use among firms with 10+ staff is 4.72% in Greece and 14.33% in the EU.

source (opens in a new tab)

Enterprises with 10+ staff. source (opens in a new tab)

The invisible problem

Breakdowns aren't random. Nobody was listening.

Friday isn't to blame. The machine runs flat out all week, and on Friday night the technician who listens to it with a screwdriver to their ear has gone home. A bearing doesn't fail without warning. It heats up, vibrates, draws more current, days before. Nobody happens to be looking.

The stop costs more than the repair: lost output, Saturday overtime, a part sent by express courier, a late delivery to a customer who scores you on “on time, in full”.

Today's alternative is calendar maintenance: change the bearing every six months, whether it needs it or not. You pay twice. For parts thrown away half-used, and for failures that arrive before the calendar does.

How it works on your floor

It listens, understands, tells someone.

1 What you already have

Sensors where it hurts

Wireless vibration and temperature sensors on the 3–⁠5 machines that stop the line: motors, pumps, fans, gearboxes, compressors. Current transformers for the motor, plus whatever the PLC already records: pressure, oil temperature, running hours. The sensors fix on with a magnet or adhesive, no cabling.

vibration · temperature · current · OPC UA · Modbus · MQTT
2 The model

The model knows what healthy sounds like

In the first weeks it learns each machine's normal signature, by load and by product. Then it looks for departures: vibration rising in mm/s RMS, frequencies that match bearing wear or misalignment, temperature climbing for no reason. It doesn't need a history of dozens of failures to get started.

anomaly detection · vibration spectrum
3 The decision, inside your system

A work order, not a siren

When the departure persists, a draft work order opens in your maintenance system or ERP, with the chart, the likely cause and the part. Your maintenance lead decides when the machine stops, inside the production plan. Nothing shuts itself down.

“Compressor A2, motor bearing: vibration 3.4 mm/s and rising, outer-race wear pattern. Suggested bearing change at the next shift change. In stock: yes.”

work order · ERP · maintenance system · Viber

Three-step diagram: sensors on a motor and pump, a chart where the live reading leaves the normal operating band, and a work order on a tablet with the likely cause and the part.

Hardware: vibration and temperature sensors and one gateway per area. Supplied and fitted by a hardware partner or the supplier you already trust. We take care of the data, the models, the dashboards and the alerts.

What we measure

Five numbers. None of them new.

Fig. 1 — A motor, a pump, a sensor on the bearing. It hears what the night shift can't.

Your maintenance lead already knows them, roughly. We measure them exactly, before and after.

Unplanned downtime hours a month
The formula Σ duration of stops that weren't scheduled
MTBF · mean time between failures
with / without the experienced technician on shift
The formula running hours ÷ number of failures
MTTR · mean time to repair
with / without the experienced technician on shift
The formula total repair time ÷ number of repairs
Maintenance cost per tonne (or unit)
The formula (labour + parts + outside contractors) ÷ tonnes produced
Availability (the first term of OEE)
The formula run time ÷ planned time · OEE = Availability × Performance × Quality

How the benefit is worked out

Benefit = downtime hours avoided × cost per downtime hour + calendar-based part changes no longer needed − running cost

Payback (months) = pilot cost ÷ (monthly benefit − monthly running cost)

Your cost per downtime hour is yours: lost margin, overtime, penalties, emergency freight. We work it out from your own figures, not from other markets' averages.

The pilot

8–⁠12 weeks. 3–⁠5 critical machines.

If the target isn't met, the report says so, with the same numbers we would have shown if it had been.

NTX-PL-PM8–⁠12 weeks

Pilot charter

Scope: We choose together the machines that stop the line when they stop, have no standby, and take expensive or slow-to-arrive parts.
Duration: 8–⁠12 weeks
Phases: Weeks 1–⁠2Machine criticality from the breakdown log, sensors fitted, PLC connected.
Weeks 3–⁠6The model learns normal running. A dashboard for maintenance.
Weeks 7–⁠12Live alerts and work orders. Measurement against the baseline. Decision.
Target: Example target: every confirmed departure (wear the technician found on opening the machine, a known fault, a controlled test) is caught before failure, with false alarms below the weekly limit your maintenance lead sets.¹
Cost: It starts with a Plant X-Ray, from €4,900, credited in full against the pilot.Sensors are priced separately.

Signed before any code is written.

¹ In 12 weeks, nothing may break. So the target doesn't say “predict a failure”. It says what was caught, what was confirmed and how many alarms were false. Downtime hours are judged over a year.

Evidence

What the numbers say, and who's saying them.

≈2,000 hours
of unplanned downtime and 53 process interruptions avoided over three years at steelmaker BlueScope, with predictive maintenance (Siemens, 2022).
source (opens in a new tab)
27 hours
of unplanned downtime a month, on average, at a large plant.
source (opens in a new tab)
−50% · −40%
in unplanned downtime and maintenance cost respectively: Siemens' own claim for customers of its Senseye platform.
source (opens in a new tab)
69% · 3–⁠4%
of Greek SMEs say they need predictive maintenance. Only 3–⁠4% use it.
source (opens in a new tab)

Our own measured project

We haven't published a measured predictive-maintenance project yet. The first one could be in your plant.

Most of the numbers on this page come from vendors, and they are their best cases. We show them so you know the range, not as a promise.

Honest limits

It doesn't predict everything. We say so first.

Show one of the model's mistakes and how we handle it

One of its mistakes, in fullMonday 09:30

What the model said: The model said: “Conveyor C1 gearbox, vibration out of pattern.”
What actually happened: In fact the conveyor was running a new, heavier product at a higher speed. The machine was healthy.
What happened next: The maintenance lead tagged it “operating change”. The model now keeps a separate pattern for each product. The false alarm was counted in that week's report.

What it needs from you

Machines with moving parts and a steady running pattern: motors, pumps, fans, gearboxes, compressors.
The breakdown log, even if it's a notebook. We read it.
Access to the machines to fit sensors, under your safety rules.
A maintenance lead who answers the alerts. Without one, an alarm is just noise.

What it won't do

It can't predict failures that give no warning: electronics that die instantly, accidents, operator error.
It doesn't give a failure date to the day. It gives a warning and a window.
It doesn't scrap calendar maintenance overnight. It stretches it out where the data justifies it.
It doesn't touch the machine. It stops nothing on its own.

Play

The Glass Factory

One week, two endings.

A plant with five machines and data that runs live: one week in two minutes. Choose how it's maintained.

Fully playable by keyboard.

Skip the demo: read how it works

Wed 14:20 · Vibration leaves its pattern. Alert.
Thu 06:00 · A 50-minute planned stop. Bearing from the shelf.
Fri 22:40 · The line is running. The technician is asleep.

The demo didn't load. The text below says what you would have seen.

In the Glass Factory, a plant with five machines, vibration on a compressor bearing starts to climb on Wednesday afternoon. Without prediction, nobody sees it: the compressor stops on Friday at 22:40, the bearing isn't in stock, and line 2 restarts at Monday lunchtime. With predictive maintenance, the model gets one wrong first: on Monday it raises an alarm on a conveyor gearbox that was simply running a heavier product, and the maintenance lead logs it as false. On Wednesday afternoon it correctly flags the compressor, and the bearing is changed in a 50-minute planned stop on Thursday at 06:00. On Friday night the line runs normally.

What you just saw

Mechanism: Detecting departures from each machine's normal signature, with a warning before failure.
Why it works: Most mechanical failures give signs days in advance. The problem isn't that the signs aren't there. It's that nobody is looking at that moment.
Data it needs: Vibration, temperature and current from the critical machines; the breakdown log; the production plan.
Honest range: How many hours your plant saves depends on how often its machines fail and what each stop costs. We measure it over a year. We don't promise it in advance.

Which machine ruins your Fridays?

Same logic, another trade

Every trade has its own Friday night.

Same mechanism: sensors, a normal-running pattern, an alert with time to act. Six machines that can't afford to stop.

Use cases: the same mechanism at work in other trades.

Cotton gin, Karditsa