Predictive Maintenance in a Malaysian Factory: What It Saves, What It Costs, and Which Machines to Start With

A production manager at a Shah Alam plant described the problem to us in one sentence: "We service everything on schedule and things still break."

That is the honest starting point for most Malaysian factories, and it is not a failure of discipline. It is what a time-based maintenance programme does. You service the machine because the calendar says so, not because the machine says so. A bearing that was going to fail in week seven fails in week seven regardless of whether you changed the oil in week four.

Predictive maintenance is the alternative, and it is worth understanding properly before anyone sells you a dashboard. This guide covers what it genuinely saves, what it costs including the parts vendors leave out, and which machines to start with, which is the decision that determines whether the project pays back.

What predictive maintenance actually is

The distinction is simple and it is the whole idea: predictive maintenance is condition-based, preventive maintenance is time-based.

The US Department of Energy, in its O&M Best Practices Guide, defines it as measurements that detect "the onset of system degradation… allowing causal stressors to be eliminated or controlled prior to any significant deterioration in the component physical state."

Their analogy is the clearest one available. Most people change the oil in a car every 5,000 kilometres because that is the interval — a preventive task. If instead you analysed the oil periodically and changed it when the analysis said it was finished, you might get to 10,000 kilometres. Same machine, same goal, entirely different trigger.

On a factory floor the measurements are vibration, temperature, lubricant condition, current draw, acoustic signature. The AI part, and this is where it earns its keep, is learning what normal looks like for your machine, so a deviation gets flagged before a human on rounds would notice it.

What it saves, with the number sourced

Here is where you should be sceptical of everyone, including us.

You will see claims of 30%, 40%, 50% downtime reduction on vendor websites. Ask where the figure comes from. The most widely-cited independent source is the Department of Energy guide above, and its number is more modest and better documented:

Estimated 8% to 12% cost savings over preventive maintenance program.

Read what that is measured against. Eight to twelve percent over a preventive programme — not over doing nothing. If your factory is currently running mostly reactive maintenance, you are on a different rung entirely, and the same guide gives the whole ladder in cost per horsepower per year:

• Reactive — run it until it breaks: $18/hp/yr
• Preventive — service on a schedule: $13/hp/yr
• Predictive — service on condition: $9/hp/yr
• Reliability-centred — predictive plus root-cause analysis and redesign: $6/hp/yr

Those are US federal facility figures, not Malaysian quotations, and the horsepower-per-year convention is theirs. Use them for the shape of the argument: the biggest single jump is getting off reactive maintenance, and predictive is a further step down rather than a magic one.

One more line from the same source, which is the honest version of the downtime claim: a well-run predictive programme "will all but eliminate catastrophic equipment failures." That is a statement about the kind of failure you stop having, not a percentage.

The costs vendors leave out

The Department of Energy lists disadvantages alongside the advantages. Most vendor pages reproduce the advantages and quietly drop this half:

• Increased investment in diagnostic equipment.
• Increased investment in staff training.
• Savings potential not readily seen by management.

That third one is the project killer, and it deserves more attention than it gets. A failure that does not happen produces no invoice. There is no line item, no dramatic recovery, no visible win — just a quarter where nothing went wrong, which looks identical to a quarter where you were lucky. Programmes get cancelled in year two for exactly this reason.

The practical defence is to measure the right thing from day one. Not "savings", which are invisible, but the leading indicators the same guide benchmarks: equipment availability above 95%, and emergency maintenance below 10% of total maintenance hours. If emergency work is falling as a share of your hours, the programme is working, and that is a number your finance director can see.

Which machines to start with

This is the decision that determines payback, and it is where most budgets get wasted.

The instinct is to instrument the most expensive machine. That is often wrong. The Department of Energy's reliability-centred maintenance framework sorts equipment by which approach suits it, and the split is about failure behaviour rather than value:

Good candidates for predictive — equipment with random failure patterns, critical equipment, and equipment not subject to predictable wear. These are machines where a schedule cannot help you, because the failure does not arrive on a timetable.

Better left on preventive — equipment subject to wear with known failure patterns, and anything where the manufacturer's service interval genuinely reflects how the part degrades. Putting sensors on a machine with a known wear curve buys you alerts that scheduled servicing would have handled anyway.

Fine to run to failure — small parts and equipment, non-critical equipment, equipment unlikely to fail, and redundant systems. Not everything deserves monitoring, and pretending otherwise is how a pilot turns into a budget conversation.

So the first-line question is not "what is our most important machine?" It is: which machine fails without warning, hurts when it does, and does not follow a wear curve we already understand? That machine is your pilot.

What a Malaysian deployment actually involves

The reassuring part is that you do not replace anything. Existing machines are instrumented with sensors and edge gateways — no production-line replacement — which is what makes it possible to prove the case on one line before committing further.

On our own numbers: a pilot covering one production line typically starts from RM50,000 to RM150,000, and a facility-wide programme runs from RM200,000 upward. A focused pilot can be deployed in four to eight weeks; a full transformation across multiple lines is a three-to-nine-month piece of work. We set that out in more detail on our smart manufacturing service page.

Older machines are frequently better pilot candidates than new ones, for the unglamorous reason that they fail more and therefore give the model more to learn from.

The part that decides whether it works

An alert nobody acts on is worse than no alert, because it costs money and teaches the team to ignore the screen.

Before the sensors go on, settle three things in writing:

• Who receives the alert, by name or by rotation — not "the maintenance team".
• What threshold triggers a work order rather than a note, so the shift lead is not making that judgement at 2am.
• What happens when the model is wrong, because it will be. A first-month false-positive rate is normal, and the plan for handling it is the difference between a programme people trust and a programme people mute.

This is the same pattern that decides whether any AI system survives contact with an operation, and we wrote about it more generally in what an AI agent actually costs to build and run — the monitoring and the exception queue are the ongoing cost, not the model.

When not to do this

Worth saying, because the honest answer is sometimes no.

If your machines follow predictable wear curves and your scheduled servicing is genuinely being completed, predictive maintenance is a marginal 8–12% improvement on something already working: real, but not urgent. If your maintenance is mostly reactive, the bigger and cheaper win is getting a preventive programme running properly first. That is the $18 to $13 step, and it needs discipline rather than sensors. And if nobody in the plant owns the alerts, buy nothing until somebody does.

The factories that get the most from predictive maintenance are the ones with a specific, expensive, unpredictable failure they can already name. If you can name yours, that is the whole business case.

If the recurring problem is defects rather than breakdowns, AI quality inspection covers the other half — and why the camera is rarely the part that is missing.