What is predictive maintenance, and how does it differ from preventive?

Four maintenance strategies

Maintenance strategies differ by what triggers the work. Reactive maintenance repairs equipment after it breaks. Preventive maintenance performs tasks on a calendar or running-hour schedule, regardless of the machine's state. Condition-based maintenance acts when a measurement, such as vibration or temperature, crosses a set limit. Predictive maintenance follows the trend of the same measurements and estimates when that limit will be reached. The steps are usually taken in order of data maturity. Condition-based and predictive work need sensor data, a defined limit and disciplined records. Without those, starting with preventive maintenance while collecting records is usually the more sound first step.

Worked example: a pump bearing (hypothetical)

Suppose a pump runs 6 000 hours a year, and its bearing is replaced every 3 000 running hours. Over two years, or 12 000 hours, that rule gives 12 000 / 3 000 = 4 replacements. Suppose vibration trending shows the bearing reaching the alarm limit at about 4 500 hours. The team plans the replacement at 4 000 hours to keep a margin. Over the same two years the count is 12 000 / 4 000 = 3 replacements. One replacement and one planned stop are avoided. These numbers show only the arithmetic of the plan. They do not prove the trend forecast is correct, so every new measurement should be checked against the plan.

Limits and common mistakes

Predictive maintenance can only see what the sensors measure. A fault whose signature falls outside the measured frequency band, such as a lubrication problem or misalignment, may appear late in the trend or not at all. The most common mistake is to start with a model and leave out the decision rule: no threshold, no owner, no response time. An alarm without a defined action adds work for the maintenance crew. Another mistake is to claim accuracy without failure history, since no past failures means the forecast cannot be tested. Each forecast needs a validation period, a comparison with maintenance records and a log of the operating conditions.

Follow-up questions