Predictive maintenance ROI: how to calculate it from your own plant's data

What predictive maintenance ROI means

Predictive maintenance ROI (return on investment) is the ratio between the net benefit of a predictive maintenance program and the cost of running it over the same period. The question it answers is simple: for every rupiah spent on sensors, software, and team time, how much loss did not happen? The answer is less simple, because most of the benefit consists of events that never occurred, so it has to be estimated with care.

If the underlying concepts are new, the articles on predictive maintenance and condition monitoring explain how they work. The focus here is the economics: how to build numbers that hold up in front of management and the finance department.

One principle needs to be agreed at the outset: ROI is calculated per asset group, not for the whole plant at once. The reason lies in how predictive maintenance works. Condition monitoring only helps with failure modes that give a warning before they occur, and it only pays when the consequences of that failure are expensive enough. Moubray states this condition within the Reliability-Centred Maintenance framework: an on-condition task is worth doing if, over a period of time, it costs less than the operational consequences plus the cost of repairing the failure it targets^[1]. A plant-wide average mixes assets that are well worth monitoring with assets that are cheaper to run to failure, and the result explains nothing.

Choosing which assets to include

Because ROI belongs to an asset group, the first step is to choose that group for reasons that can be traced. ISO 17359 places criticality assessment and failure mode analysis at the start of a condition monitoring program, before techniques and measurement points are chosen^[2]. In practice, three questions help screen the candidates.

• Consequence of failure. What happens if this asset stops without warning? A transfer pump that stops the whole filling line is in a different class from a ventilation fan with a standby unit. Consequences include lost production, safety, quality, and repair cost.
• Frequency of failure. How often has this asset failed in the last one or two years? The data is in the work order history, and how to read it as MTBF is covered in the article on how to calculate MTBF and MTTR. An asset that has not failed in five years leaves little room for benefit, however critical it is.
• Detectability. Does the failure mode produce a measurable symptom, such as a change in vibration, temperature, or motor current? Bearing damage and misalignment can generally be detected; a component broken by an external impact generally cannot.

The third question carries an extra condition that is often missed: the P-F interval. The time between the first detectable symptom and the loss of function must be long enough and consistent enough for the chosen technique to leave time to plan the repair^[1]. A monthly manual vibration check is of no use for a failure mode that develops in two weeks. For a failure mode like that, the ROI calculation must assume closer monitoring or leave it out altogether.

The result of this screening is usually one small group of critical rotating assets on one line, such as drive motors, pumps, blowers, and gearboxes. A small group like this is easier to calculate, easier to monitor, and easier to check afterward.

The benefit side and how to measure it from your own records

Every benefit of predictive maintenance rests on one mechanism: a failure detected early can be repaired as planned work rather than as an emergency. The difference between those two ways of working is what gets counted. Because a failure that is caught never happens, every "avoided" number is an estimate. That estimate can only be defended if there is a baseline: a record of events and costs in the period before the program starts, ideally twelve months or more so that production seasons and overhaul schedules are represented.

Unplanned downtime avoided

This is usually the largest item. From the downtime records, take the unplanned stops on the selected asset group, then separate those caused by failure modes the planned technique can detect. For each event, record how long the line was down, from the moment the machine stopped until production was back to normal. If downtime reason codes are not yet in order, the article on machine downtime explains how to set them up. The benefit is the stop hours avoided multiplied by the downtime cost per hour. A planned repair often takes production time as well, so what counts is the difference in hours, not every hour of the emergency event.

Emergency repair premium

An emergency repair is almost always more expensive than the same job done as planned work. The difference comes from technician overtime on the night shift, express-shipped parts, contractor call-outs outside working hours, and short-notice rental of lifting equipment. Compare the two from work order history and invoices: the average cost of an emergency repair for a given failure mode and the cost of equivalent work done on schedule. That difference is the emergency repair premium. In the calculation, pick one method and stay with it. Either count the full emergency cost as a benefit and put the planned repair cost on the cost side, or count only the premium and leave the planned repair cost out. Mixing the two either counts the planned repair twice or does not count it at all.

Secondary damage avoided

A failure allowed to run to the end often damages the parts around it. A seized bearing can score the shaft and damage the bearing housing, while a coupling that comes apart can damage the motor mounting. Action at an early stage usually means replacing only the worn component, so that further damage can be avoided^[1]. Measure it from past repair reports: how many times failures in this asset group caused secondary damage, and what the extra cost was. Divide that total by the number of detectable failures to get an average per event.

Spare parts, inventory, and energy

Two other items are often mentioned but are best counted with care. When asset condition is monitored, some spare parts can be ordered once symptoms appear, so certain safety stocks can be reviewed. But stock only falls after purchasing decisions change, and for long lead time parts that may never be safe to do. Energy is relevant to some failure modes, for example misalignment or poor lubrication that can add load to a motor, or compressed air leaks found by ultrasonic inspection. If the difference in energy use is not measured directly before and after the repair, it is more honest to leave this item at zero in the first proposal.

The cost side of predictive maintenance: more than the sensor price

A proposal that lists only the price of the sensors will be questioned by finance, and rightly so. Jardine and colleagues describe a condition-based maintenance program as three interdependent steps: data acquisition, data processing, and maintenance decision-making^[3]. Each step has a cost, and the third is the one most often forgotten. The cost items to include are as follows.

• Sensors and gateways. Vibration and temperature sensors, gateways, power supplies, and the network that carries the data. This is an initial investment paid once.
• Software subscription. The monthly or annual fee for the monitoring and analytics platform, usually based on the number of assets or measurement points. The components of maintenance software pricing are covered in more detail in the article on CMMS software pricing.
• Installation and commissioning. Fitting sensor mounts, running cables or setting up the wireless network, and any production time used during installation.
• Analysis time. Someone has to read the trends, assess alarms, and decide on action, whether an in-house engineer or an outside analyst. Count the hours per week and multiply by the cost per hour.
• Training. Technicians and engineers need to understand what an alarm means and how to follow it up. Mobley stresses that a predictive maintenance program depends on trained people, not only on its instruments^[4].
• Cost of acting on findings. Planned repairs still need parts and labor hours, and every alarm that turns out to need no repair still takes inspection time. This item often goes missing from proposals, yet without action no benefit is realized.

The initial investment (sensors, gateways, installation, and training) is best kept separate from recurring costs (subscription, analysis time, and acting on findings). The separation matters because the two kinds of cost are treated differently in the formulas that follow.

Formulas for ROI, payback period, and downtime cost per hour

The three formulas below are enough for most proposals. The first sets the value of the largest benefit item, so its number should be agreed with finance before any other number is built.

ROI answers whether the program returns more than it costs, while the payback period answers how long it takes for the initial capital to come back. Management usually asks both. For supporting indicators such as maintenance cost or availability, EN 15341 provides standard definitions of maintenance key performance indicators, so the numbers before and after the program can be measured the same way^[5].

A worked example for one asset group

The figures below are illustrative, not data from any particular plant and not a forecast. A bottle filling line has 10 motors and pumps rated as critical. Over a 12-month baseline, this asset group had 8 unplanned stops. Six were caused by failure modes that vibration and temperature monitoring can detect: bearing damage, misalignment, and poor lubrication. The other two, sudden electrical faults, are left out of the calculation.

The other assumptions come from the same records. Each emergency event stopped the line for 6 hours on average, while a planned repair of the same finding is expected to still take 1 hour of production time, so the hours avoided per event are 5. The line produces 3,000 bottles per hour at a contribution margin of IDR 2,500 per bottle, or IDR 7,500,000 per hour. Twelve operators stand idle at IDR 50,000 per person per hour, or IDR 600,000 per hour. Recovery costs over the baseline were IDR 32,400,000 for 36 stop hours, or IDR 900,000 per hour. The downtime cost per hour is IDR 7,500,000 + IDR 600,000 + IDR 900,000 = IDR 9,000,000.

An emergency repair cost IDR 14,000,000 per event on average, including overtime, express parts, and contractor call-outs, while equivalent scheduled work cost IDR 6,000,000. Two of the six failures also damaged the shaft and bearing housing at a total extra cost of IDR 42,000,000, or an average of IDR 7,000,000 per detectable failure. The team expects the program to catch 4 of those 6 failures early enough. The initial investment is IDR 84,000,000 (sensors and gateways IDR 60,000,000, installation IDR 16,000,000, and training IDR 8,000,000), spread evenly over three years.

Two things stand out in this table. First, downtime accounts for IDR 180 million of the IDR 264 million benefit, so the downtime cost per hour drives the result more than any other number and is the one most worth checking with finance. Second, the cost of acting on findings (planned repairs and alarm checks) comes to IDR 32 million, more than a quarter of the annual cost. Dropping that item would raise ROI on paper without changing anything on the plant floor.

Sensitivity: what if only half are caught

The figure "4 of 6 caught" is the weakest assumption in the calculation above. Detection depends on sensor mounting, measurement frequency, alarm limits, and how quickly the team responds. A prognosis of a component's remaining life always carries uncertainty, and ISO 13381-1 asks that the confidence level of a prognosis be stated along with the result^[6]. The same attitude belongs in an ROI calculation: show how the result changes if that assumption misses.

With half the catches, the program still produces a net benefit, but ROI falls from 120.0% to 22.2% and payback stretches from about 5.9 months to about 19.4 months. If only one failure is caught, the program loses money. The break-even point can be calculated directly: fixed costs of IDR 96 million divided by the net benefit per catch, which is IDR 66 million less IDR 6 million of planned repair cost. The result is 96 ÷ 60 = 1.6 failures per year.

Steps to run a pilot that produces its own ROI evidence

For many plants, implementing predictive maintenance is easiest to start with a pilot designed from the outset to produce evidence. Mobley recommends that a predictive maintenance program have measurable goals from the start, so that its benefits can be compared with conditions before the program^[4]. The following sequence can be used.

1. 1Lock the baseline. Collect the last 12 months of data for the pilot asset group: number and duration of unplanned stops, their causes, repair costs, and secondary damage. Agree the downtime cost per hour with finance before the pilot starts.
2. 2Choose the pilot assets. Use the criteria of consequence, frequency, and detectability. One small group on one line is easier to follow than assets scattered across the plant.
3. 3Make a first estimate. Put the baseline into the formulas above, or use the ROI simulator to try several scenarios, then prepare the base case, the conservative case, and the break-even point.
4. 4Agree success criteria before starting. Write down what counts as a valid finding, for example an alarm followed by an inspection that shows real damage, documented with photos or the removed component. Also write down the number that must be reached for the program to be extended.
5. 5Record every alarm and its action. For each alarm, record the work order, the finding, the planned repair cost, and the estimated consequence had it been left. Alarms that produce no finding are still recorded together with their inspection time.
6. 6Review after 3 to 6 months. Compare valid findings with the baseline failure rate for a period of the same length, recalculate ROI with real numbers, then decide whether the pilot is extended, adjusted, or stopped.

Common mistakes and limits of the calculation

• Counting every alarm as an avoided failure. An alarm only becomes a benefit when the inspection finds real damage that would have developed into a failure. Alarms caused by load changes, loose sensors, or limits set too tight are a cost, not a benefit.
• Ignoring the cost of acting on findings. Planned repairs still need parts, labor hours, and sometimes production time. Counting the full emergency cost as a benefit without including the planned repair cost makes ROI look larger than it is.
• Picking non-critical assets. Assets that are cheap, have a standby, or rarely fail produce no measurable benefit, however sophisticated the monitoring. A pilot on assets like these can end with the false conclusion that predictive maintenance does not work.
• A pilot that is too short. If the baseline records six failures a year, a one-month pilot may not see a single one. The pilot length should match the expected failure frequency, not the budget cycle.
• Using numbers from outside the plant. Savings percentages from brochures or industry publications do not carry over to a particular line and its assets. The only benefit that holds up is one that can be traced to your own plant records.
• Treating ROI as a promise. A tidy calculation is still an estimate. Failures that develop too quickly or are not measured by the installed sensors can still occur, and the benefit figures should be updated once pilot data is available.

Conclusion

A credible predictive maintenance ROI does not start from a savings figure but from your own plant records: how often a given asset group stopped, for how long, and at what cost. From that baseline, benefits are estimated per asset group and per detectable failure mode, while costs are counted in full, including analysis time and the cost of acting on findings.

A strong proposal presents the base case, the conservative case, and the break-even point openly, then proposes a pilot with success criteria agreed before it starts. After three to six months, the estimates can be replaced with real numbers, and the decision to extend the program is made on evidence, not on promises.