Predictive maintenance, where it pays and where it does not

Wireless vibration sensors, battery-powered and connected to a gateway, have become cheap enough to fit the budget of an average plant. With the price came the sales pitch: AI platforms that “predict failures” across the whole plant. Some of them do work. For a plant, two questions remain: on which equipment does the effort pay back, and who looks after the system once the vendor leaves.

The P-F interval

The basic idea comes from the report “Reliability-Centered Maintenance” by Nowlan and Heap, written in 1978 for the US Department of Defense on the basis of civil aviation experience. Many faults do not appear suddenly. There is a point P, the potential failure, from which degradation can be detected, and a point F, the functional failure, when the equipment no longer does its job. The interval between them is the window in which you can plan the intervention.

Two practical consequences follow. Measurement must be taken more often than the P-F interval lasts, otherwise the fault slips between two readings. And the interval must be long enough to order the part, negotiate a stoppage with production and carry out the work. A bearing that degrades over several weeks is a good candidate. A fault that goes from first sign to breakdown in a few minutes cannot be caught with a monthly route, however good the instrument.

What can be detected

Techniques with a long track record in industry cover several families of faults:

• Vibration analysis catches bearing wear, imbalance, misalignment, mechanical looseness and some gear faults. On rotating machines at constant speed it is the most mature method. ISO 20816-1 gives the general framework for measuring and evaluating vibration on machines, and the other parts of the series cover specific machine types.
• Infrared thermography shows loose or oxidised electrical connections, unbalanced phases, heated fuses and contactors in cabinets, as well as hot bearings or damaged insulation.
• Oil analysis shows contamination with water or dust, lubricant degradation and wear particles, useful on large gearboxes, hydraulic systems and compressors.
• Ultrasound detects leaks of compressed air, steam and vacuum, partial discharges in electrical equipment and, when used for lubrication, shows when a bearing really needs grease.

What cannot be seen

Many faults have no useful P-F interval or give no signal that a sensor can read. Electronics (control boards, power supplies, drivers) usually fail without warning. So do overload fractures, collisions, operating errors and software faults. A cable that breaks inside a cable carrier or an inductive sensor that gets hit gives no early notice.

Condition monitoring does not help with these. What helps is redundancy, spares on the shelf, visual inspections or, sometimes, accepting the failure and repairing quickly. If an offer promises to “predict any failure”, ask for concrete examples of failure modes on equipment like yours.

Consequence decides

The fact that a fault can be detected does not mean it is worth monitoring. RCM thinking starts from the consequence of failure: what happens if the machine stops unplanned. A pump with a parallel standby, swapped in an hour, can run until it fails without anyone losing money. A single exhaust fan on a paint line, a compressor with no backup or a gearbox with a lead time of months is a different case.

This is why predictive maintenance comes after a criticality ranking, not before it. Take the equipment where a stop costs a lot or affects safety, check whether its dominant failure modes are among the detectable ones and only then choose the technique. ISO 17359, the general guide for condition monitoring programmes, follows the same order: equipment audit, then reliability and criticality audit, and only then the measurement method.

The cost the quote does not show

Hardware is the easy part. The real cost is the person who looks at the alarms and someone who acts on them. A system with 50 sensors generates alerts every week. Some are false (a detached sensor, a change of operating regime, a machine stopped for cleaning), others are real and need a decision: schedule the intervention or keep watching.

If nobody in the plant has time allocated for this and knows how to read a spectrum, the alarms are ignored within months, while the platform subscription keeps being paid. The options are to train an in-house person, pay an external analysis service or combine the two. Whichever you choose, ask in advance who receives the alert, how fast they respond and how the recommendation reaches the maintenance plan.

Early detection without a timely planned intervention changes nothing. If the team is stuck in emergencies anyway and cannot do basic PM, predictive maintenance only adds a list of problems known in advance.

How to start

A reasonable start, with little money and little risk:

1. Choose five to ten critical rotating machines with a history of mechanical faults.
2. Start with manual measurements on a route: a portable vibration collector, a thermal camera, an ultrasound detector. A technician visits the marked measurement points monthly.
3. Build baseline values for each machine and follow the trend. Change against a machine’s own history often says more than an absolute value.
4. Record every finding, what was done and what was found at teardown. This is how you learn whether the method catches faults in time.
5. Fit permanent sensors where the monthly route is too infrequent for the P-F interval, where access is dangerous or difficult, or where the consequence of a stop justifies continuous monitoring.

What remains unclear is how much AI models are worth in practice compared with classic threshold and trend alarms, especially on machines with variable operating regimes. It is worth asking the vendor for a pilot on a few machines, with success criteria set in advance: how many real faults caught, how long beforehand, how many false alarms.

If you have pilot results in predictive maintenance, good or bad, you can discuss them in the Maintenance & Reliability group.