In many factories the organisation's most valuable data asset is the maintenance log: the book in which every failure, every part replacement and every unplanned stop is recorded, and which usually stays in the maintenance supervisor's desk drawer. Organised properly, that same history shows clearly what signs each machine gave before it failed.

Time-based maintenance runs on the calendar: replace every thousand hours. It is simple, but it carries two costs; a part with life left in it gets replaced, and a part that wore out early takes the line down. Condition-based maintenance looks at the actual behaviour of the machine instead of the calendar: vibration, temperature, current draw and unusual noise.

Three areas that pay off first

• Visual quality inspection: detecting surface, dimensional or assembly defects on the line with an industrial camera.
• Predictive maintenance: warning before a failure, based on trends in sensor data.
• Process parameter tuning: finding the combination of temperature, pressure and speed that minimises scrap.

Why visual inspection is usually the first win

The data for visual inspection can be created in a few weeks: it is enough to photograph good and defective parts systematically. Failure data, by contrast, cannot be created; you have to wait for it to happen. That is why a project starting with inspection reaches a demonstrable result sooner, and that result buys management support for the next steps. Success here depends less on the model than on lighting, angle and camera stability.

The rare-event problem in predictive maintenance

A critical motor may fail once a year, which means that even with continuous logging the number of failure examples is small and the model has little to learn from. There are two practical routes: first, flag deviation from the machine's normal behaviour rather than predicting the failure itself; second, start recording every event precisely today, even if you build the model next year. Factories that got their logging in order earlier have more options later.

Why it matters for Iranian SMEs

The cost of a stopped line in Iran is not only lost output; sourcing a spare part can take weeks, and that wait turns a short stop into a long one. A warning that arrives a few days early creates the chance to order the part before the failure, and its value is not comparable to the same warning in a market where the part arrives tomorrow. On the other side, many Iranian production lines already carry the necessary sensors and simply do not record their data.

A 90-day route to start

• Days 1–30: choose one critical machine or station, the one whose stoppage halts the whole line. Move its repair history from the notebook into a table and set up automatic sensor logging.
• Days 31–60: build simple trend-based alerts rather than a complex model; in many cases a gradual change in vibration or temperature is plainly visible before a failure. Enable the alerts for the maintenance team only at first.
• Days 61–90: record every alert and the outcome of its inspection; that table is the training data for the next stage. Then compare unplanned downtime on that machine with the period before you started.

Frequent mistakes

• Instrumenting every machine before proving that data from one machine leads to a better decision.
• Setting alert thresholds so low that several pointless alerts appear daily; the maintenance team quickly learns to ignore them.
• Disregarding operator knowledge; someone who has run that machine for ten years recognises signs no sensor records.
• Running visual inspection under variable light and angle, so the model reacts to shadow differences rather than to the defect.

Three actions for this quarter

• Identify the machine with the most downtime over the past twelve months and start there.
• Digitise the maintenance log; even a spreadsheet with date, cause and duration is enough.
• Record unplanned downtime and the scrap rate as your baseline.

Frequently asked questions

• What if our machinery is old and has no sensors?
  Retrofit vibration and temperature sensors can be fitted to older equipment, and their cost is usually negligible against a single long stoppage.
• Does visual inspection replace the human inspector?
  In practice it frees the inspector from examining every part and refocuses them on suspect items and root-cause analysis; borderline cases remain a human decision.
• How much data is needed to start?
  For visual inspection, a few hundred good and defective samples are often enough for a pilot; for predictive maintenance, consistency of logging matters more than volume.

Takeaway

In manufacturing, AI does not replace engineering knowledge; it puts that knowledge on top of data that has not been recorded until now. One critical machine, one organised maintenance log and one clear metric — unplanned downtime — make a starting point whose cost is trivial next to a single lost production day.

Glossary

• Predictive maintenance: repair based on genuine signs of wear, before a failure occurs.
• Condition-based maintenance: deciding to repair from the machine's current state rather than from a calendar.
• Unplanned downtime: the line stopping without warning, the most expensive kind of stop.
• Scrap: defective output that cannot be sold, with its material and time cost lost.
• Visual inspection: detecting defects by analysing camera images instead of examining each part by eye.
• Baseline: the recorded value of a metric before a change, so improvement can be measured.