industrial//maintenance

Maintenance is the set of actions that keep equipment able to perform its function or restore it after a failure, and its central engineering question is when to intervene: too early wastes healthy parts and production, too late pays for the failure and its consequences. Every strategy is an answer to that question with a different amount of information about the unit, and each is right somewhere.

Maintenance is the set of actions that keep equipment able to perform its function or restore it after a failure, and its central engineering question is when to intervene: too early wastes healthy parts and production, too late pays for the failure and its consequences. Every strategy is an answer to that question with a different amount of information about the unit, and each is right somewhere.

Corrective maintenance intervenes when the part breaks. It is the right answer for a cheap part whose failure has no consequence (a lamp, a non-critical fan), and the wrong one wherever a breakdown stops a line or endangers someone. Preventive maintenance replaces every NNN hours or cycles; it works when wear is clear and lives are tightly grouped, which Weibull statistics can confirm, and it wastes spares when failures are random. Condition monitoring watches an indicator (vibration, temperature, oil debris) and intervenes when it crosses a threshold; it needs an indicator that warns early enough, which is what the P-F interval measures. Predictive maintenance goes one step further and uses a prognosis of the remaining useful life to intervene when the risk of failing before the next planned window becomes excessive; it pays where failures are expensive, the indicator trends, and stops can be planned. Behind the last two sits prognostics, the prediction of one unit's remaining life.

The value lies in the decision a prognosis changes.

A perfect prognosis that changes nobody's plan is an expensive dashboard. And the bottleneck is rarely the algorithm: it is the work orders, where the field cause says fixed, so the history a model would learn from does not exist (CMMS).

One plant uses all four at once, part by part: corrective for lighting, preventive for filters with a known life, condition monitoring on hundreds of motors, predictive on the few compressors whose failure costs a week of production. Reliability-centred maintenance is the method that makes that assignment component by component, from the failure modes and their consequences (bathtub curve).

Maturity differs sharply. Condition monitoring by vibration, thermography and oil analysis is established industry; prognosis with degradation models is a niche of aircraft engines, batteries and large turbines; remaining life learned by deep networks from raw sensors is mostly research, because the jump from simulated benchmarks to real plants is the hard part.

The decision runs on the slowest clock in the system: residuals and limits run in the PLC every few milliseconds, life statistics on a server daily, and the maintenance decision in a weekly meeting, one more link in the loop with its own latency.