Predictive Maintenance: What It Needs Before It Can Predict Anything

Part of our guide hub: AI solutions and automation

Predictive maintenance is sold as a machine learning problem. In practice, for most Pakistani plants, it is a data problem followed by a discipline problem, and the machine learning is the last and smallest part.

A model that predicts bearing failure needs to have seen bearings fail, with sensor readings attached, and with a record of what was replaced and when. Most factories have none of those three.

The three levels, and where you actually are

Level 1: reactive

Fix it when it breaks. Most small and mid sized plants here. Downtime is unplanned, spares are bought urgently at a premium, and nobody knows which machines cost the most.

Level 2: preventive

Service on a schedule: every 500 hours, every quarter. Better, and it is where the majority of realistic gains sit. You over service some machines and still occasionally get surprised, but downtime becomes largely planned.

Level 3: predictive

Sensors and models telling you a specific machine will likely fail in the coming weeks. Real, valuable, and only worth attempting once level 2 is running reliably and has been generating records for a year or more.

The honest advice: if you are at level 1, the highest return project is not predictive maintenance. It is a maintenance management system with a work order history. That is far cheaper and delivers most of the available benefit.

What predictive maintenance actually requires

• Sensors on the right machines. Vibration, temperature, current draw, pressure. Not every machine, the critical ones.
• Continuous collection and storage, which means networking on the plant floor and somewhere to keep the data.
• Failure history with labels. The essential ingredient. You need records of what failed, when, and what was done, matched to the sensor data at the time. Without failures in the data, there is nothing to learn from.
• Enough time. Typically twelve to eighteen months of data including several real failures before a model is worth trusting.
• Someone who acts on the alert. A prediction nobody schedules work against changes nothing.

That third point is why most projects stall. A plant with two years of handwritten maintenance registers cannot train a model, and digitising those registers is the real first project.

Where it genuinely pays back

The economics work when a failure is expensive and a sensor is cheap. Concretely:

• Rotating equipment: motors, pumps, compressors, fans. Vibration monitoring on these is mature, well understood, and does not require bespoke machine learning.
• Machines where failure stops the line. A bottleneck machine going down at 2am costs a shift. A redundant one does not.
• Long lead time spares. If the replacement part takes six weeks to import, knowing three weeks early is worth a great deal.
• Energy intensive equipment, where degradation shows up as rising consumption before it shows as failure. Often the fastest payback and the easiest to measure.

Where it rarely pays: cheap machines, redundant machines, equipment that fails gracefully, and anything where the spare sits in your own store.

A realistic sequence

1. Digitise maintenance records first. Work orders, what failed, what was replaced, how long it took, what it cost. Six to ten weeks of work and it improves operations immediately whether or not you ever build a model.
2. Rank machines by cost of downtime. Usually three to five machines account for most of the loss. Everything else is a distraction.
3. Instrument those machines only. Sensors on five critical assets, not fifty.
4. Start with thresholds, not machine learning. “Alert when vibration exceeds this level” catches a surprising share of failures and can be running in weeks.
5. Collect for a year while running on thresholds and preventive schedules.
6. Then model, once you have failures in the data and a team that already acts on alerts.

Most of the return arrives at steps one to four. Step six adds to it; it does not create it.

What it costs

• Maintenance management system: 800,000 to 2,500,000 PKR for a plant, or an off the shelf product per user.
• Sensors and gateways: varies widely by type and by whether the machine already has instrumentation. Budget per asset, and get a survey before believing any number.
• Data platform and threshold alerting: 1,000,000 to 3,000,000.
• Model development, once data exists: 1,500,000 and upward, and it is ongoing work rather than a one off build.

Related reading: manufacturing ERP requirements, what running an AI feature costs, and where AI genuinely absorbs operational work.

Frequently asked questions

Do we need machine learning for predictive maintenance?

Not initially. Threshold based condition monitoring on vibration and temperature catches a large share of failures on rotating equipment and can be deployed in weeks. Models add value later, on top of that foundation.

How much data do we need?

Enough to include real failures, which usually means twelve to eighteen months of continuous readings with matched maintenance records. Data without failures teaches a model nothing useful.

Can we retrofit old machines?

Yes. External vibration and temperature sensors and current monitoring work on equipment with no built in instrumentation, which describes most plants here.

What is the fastest payback in this area?

Digitising maintenance records and ranking machines by downtime cost. It is the cheapest step, it improves operations immediately, and it is the prerequisite for everything else.

Ezitech builds industrial data, monitoring and AI systems for manufacturing clients, and will tell you plainly when the first step should be simpler than the pitch. See our AI solutions or describe your most expensive downtime.