Technology · Forecasting

Predictive Maintenance: turning sensor data into maintenance you can plan

A stoppage that nobody planned costs more than one in the maintenance schedule. Predictive maintenance turns some of those unplanned stoppages into planned work.

Predictive maintenance: definition and how it differs

Predictive maintenance is a maintenance strategy that bases the timing of an intervention on the measured condition of the equipment rather than on the calendar. A model uses sensor and maintenance data to estimate the remaining useful life of a component, and the maintenance team schedules the work for a window where it costs little.

There are four approaches. Reactive maintenance means repairing after a failure. Preventive maintenance replaces parts at fixed intervals, and so either too early, while they are still intact, or too late. Condition-based maintenance relies on condition monitoring, which measures the current state and raises an alarm once a threshold is exceeded. Predictive maintenance builds on condition-based maintenance and also tells you when an abnormal reading is likely to turn into a failure.

The benefit depends on the difference between a threshold alarm and a forecast. A threshold alarm reports a condition that has already occurred. Depending on the damage, little time may be left, and an emergency repair follows. A forecast with several days' notice lets you order spare parts, schedule staff, and move the work into a low-load period, that is, a shift with few orders.

How predictive maintenance works

In technical terms, the approach is a chain of data collection, data preparation, the model, and its integration into operations, and the project can fail at any link.

StepWhat happensWhat it produces
1. Bring the data togetherSensor readings (vibration, temperature, current draw, cycle times) are linked with the fault codes and the maintenance history, including the actual failuresone data set per machine in which the failures are dated
2. Build signalsper-second readings become indicators per shift: trend, spread, outliers, and load changesfeatures that make wear visible
3. Estimate remaining useful lifeA survival analysis model estimates how likely a failure is in the coming weeks. A neural network for time series (LSTM) can detect a pattern in the sensor history that precedes a failurea failure probability per machine and period
4. Test against the calendarthe model is compared with the existing maintenance schedule on warning time and false alarmswhether the forecast is better than the fixed interval
5. Build it into maintenancethe warning arrives in the maintenance system as a recommended work order, with machine, period, and proposed maintenance windowa decision taken by a named person
6. Monitor and retrainthe hit rate is measured continuously, and the model is retrained after modifications or when new materials are introduceda model that keeps pace with modifications and new materials

Step 4 is the easiest to skip, yet without a comparison with the simple maintenance schedule nobody knows whether the model is worth the money. So ask every provider, including us, which benchmark the forecast was tested against and how many false alarms came with each justified warning. You can recognize a reputable provider by two things: they give you both answers before they talk about the model, and they tell you when the project would not pay off for you.

Step 5 then determines whether anyone acts on the forecast, because a warning needs a work order in the maintenance system and someone responsible for it.

Worked example: 48 machines, 900 hours of unplanned downtime

The following example is a model calculation for a fictitious company from our manufacturing series and contains no client figures. A metalworking business runs 48 machines in three shifts, six of them bottleneck machines.

The full model calculation in six phases
AssumptionValue
Unplanned downtime hours per year900
Follow-on cost per unplanned hour (blended rate of lost contribution margin on the bottleneck machines and idle staff at the other machines)750 euros
Unplanned downtime cost per year675,000 euros
Median warning time of the model11 days
Share of unplanned hours that becomes plannable as a result32 percent

With a median of eleven days' notice, the maintenance team can move some interventions into low-load periods. The model calculation assumes that this turns 288 of the 900 unplanned hours into planned ones and leaves 612 hours unplanned. At a follow-on cost of 750 euros per unplanned hour, the 288 hours amount to a gross benefit of 216,000 euros a year.

Every value in the table is an assumption. You can establish the hours and the hourly rate in your own plant. The plannable share, however, is something the model has yet to prove. So every calculation of this kind needs a cross-check:

ScenarioShare made plannableHours made plannableGross benefit per year
cautious20 percent180135,000 euros
model calculation32 percent288216,000 euros
favorable40 percent360270,000 euros

To see what such a forecasting model looks like in a factory, with its failure curve, maintenance alert, and pilot on one production line, go to the predictive maintenance use case in manufacturing.

For a rough calculation of your own, you need only two figures from your plant and one assumption about the plannable share:

Gross benefit per year = unplanned downtime hours × follow-on cost per hour × share made plannable

The result is a gross figure because two items that vary widely from plant to plant are deliberately left out of the formula. The first is false alarms, because every warning that is not followed by damage leads to an unnecessary intervention. The second is the cost of the planned intervention, above all its downtime hours, because the intervention still takes place, though at a cheaper time. Both belong in the evaluation before a model goes live.

To calculate the payback period in years, divide the one-off cost by what remains of one year's gross benefit after the running costs, that is, the costs of operating the model, of false alarms, and of the planned interventions. If your plant has, say, only 36 unplanned hours a year and the hourly rate and plannable share match the example, the gross benefit comes to 8,640 euros a year, and a figure of that size will not justify a project. In that case calendar-based maintenance is the more economical choice, and we will tell you so.

When predictive maintenance pays off and when it does not

A predictive approach pays off where three conditions are all met: unplanned downtime is expensive, wear shows up early in measurable values, and past failures are documented.

The third condition is the first hurdle. A model learns remaining useful life from the intervals between two failures. If the failure dates are missing from the maintenance history, or exist only as free text in a shift log, a data-driven model of remaining useful life has no foundation. The project then begins with proper fault recording, and the model follows later.

Four mistaken conclusions are easy to draw:

  • More sensors do not replace a failure history. Additional measuring points supply new signals, but not the dated failures from which a model could learn remaining useful life.
  • A pilot on a non-critical machine proves little. The benefit arises on the bottleneck machines, so that is where the pilot belongs.
  • Rare failures are not enough to forecast remaining useful life. If a machine type fails only every few years, the model lacks examples. In that case, condition monitoring with thresholds gets you further, as does anomaly detection that learns normal operation.
  • A warning without an owner comes to nothing. Before the project starts, it has to be settled who decides whether a shift is rescheduled in response to the forecast.

We advise against a project if downtime costs little, if the damage gives no advance warning or if the history is missing and cannot be built up in the foreseeable future.

Maintenance forecasting in the data science process

A project like this passes through the six phases of CRISP-DM, a process model for data science projects. Modeling is only one of them. The business question in Phase 1 determines how much of the downtime can be avoided at all. Phase 2 checks whether the failures are properly dated. Phase 3, data preparation, turns the per-second readings into the signals from which the model recognizes wear. In Phase 4 the model is built and measured against the calendar-based maintenance schedule, before Phase 5 checks whether the warning time is long enough to act on. Phase 6 brings the warning into the maintenance system and monitors whether the forecast still holds after modifications.

All six CRISP-DM phases at a glance Phase 3: data preparation Phase 6: deployment and monitoring

Frequently asked questions about predictive maintenance

What is predictive maintenance?

Predictive maintenance is maintenance driven by forecasts: a model uses sensor and maintenance data to estimate when a component is likely to fail, so that the work can be done beforehand and at a convenient time. It differs from condition monitoring in its warning time, and from calendar-based maintenance in its link to the measured condition of the equipment.

What is the difference between predictive maintenance and condition monitoring?

Condition monitoring measures the current condition and reports when a threshold is exceeded. Predictive maintenance uses the same readings but also forecasts when the abnormal reading will turn into a failure. The difference lies in the warning time, and that warning time determines whether the intervention can be planned.

What data does predictive maintenance need?

It needs continuous sensor readings such as vibration, temperature, and current draw, the fault codes from the machine controls, and a maintenance history in which past failures are recorded with their dates. The history is particularly important, because it cannot be measured after the event. Without documented failures, a data-driven model of remaining useful life has no foundation.

Which machines is predictive maintenance suitable for?

It suits machines whose failure is expensive and whose wear shows up in measured values in advance, for example bottleneck machines with rotating or thermally stressed components. The rough calculation of downtime hours, follow-on cost per hour, and plannable share shows whether a machine qualifies. For inexpensive or redundant machines, that calculation tends to favor the calendar-based maintenance schedule.

What does predictive maintenance cost and when does it pay off?

The calculation has three variables: unplanned downtime hours, follow-on cost per hour, and the share that the advance warning makes plannable. Their product is the gross benefit per year. Set against it are the one-off cost, ongoing operation, and the costs of false alarms and planned interventions. Only after deducting them can you tell whether the project pays off.

Does predictive maintenance need new sensors?

Not necessarily. If machine controls and control systems already record current draw, temperature, and cycle times, the first step is an inventory of this data and its gaps. New sensors are only worth installing once it is clear which signal is missing and on which machine it would reveal a failure earlier.

How long does it take for predictive maintenance to deliver results?

It depends on the data and the pilot. If past failures are dated and sensor readings complete, the forecast can be tested early against known failures. Without fault recording, the history has to be built up first. The pilot then shows whether the warning time holds in live operation. A reliable timescale is therefore only possible after the data check.

How does a predictive maintenance project start?

It starts with a business question and a data check. First, work out what an unplanned hour costs and on which machines such hours occur. Then check whether the failure history is sufficient for a forecast. Only then do you run a pilot on a bottleneck machine to measure the forecast against the existing maintenance schedule.

Run the numbers on your downtime with us

Two figures from your plant and a cautious assumption are enough for a first assessment of whether predictive maintenance pays off for you.

Book a free initial consultation