industrial//maintenance//predictive maintenance
Predictive maintenance is the maintenance strategy that uses a prognosis of a unit's remaining life to decide when to intervene, choosing the moment at which the risk of failing before the next opportunity becomes too expensive to carry; it sits above condition monitoring, which reacts to an indicator crossing a limit, and it pays where failures are expensive, the indicator trends and stops can be planned. A perfect remaining useful life is worth nothing if nobody changes the plan because of it.
Predictive maintenance is the maintenance strategy that uses a prognosis of a unit's remaining life to decide when to intervene, choosing the moment at which the risk of failing before the next opportunity becomes too expensive to carry; it sits above condition monitoring, which reacts to an indicator crossing a limit, and it pays where failures are expensive, the indicator trends and stops can be planned. A perfect remaining useful life is worth nothing if nobody changes the plan because of it.
The basic decision is binary. Intervene now, with an unplanned stop that costs CAC_ACA, or wait for the planned stop, where replacing the part costs CPC_PCP (little, since no production is lost), accepting the risk of a failure before it, which costs CFC_FCF. If ppp is the probability of failing before the planned stop, read from the RUL distribution, waiting costs on average pCF+(1−p)CPpC_F+(1-p)C_PpCF+(1−p)CP, and intervening pays when that exceeds CAC_ACA:
p>α∗=CA−CPCF−CPp>\alpha^{*}=\frac{C_A-C_P}{C_F-C_P}p>α∗=CF−CPCA−CP
α∗\alpha^{*}α∗ is the acceptable risk, and the costs set it, never the intuition of whoever trained the model. With round numbers for a fan bearing: 2,000 euros at the planned stop, 12,000 if the line stops four hours to do it now, 80,000 if it seizes and takes the shaft with it. Then α∗=10,000/78,000≈0.13\alpha^{*}=10{,}000/78{,}000\approx0.13α∗=10,000/78,000≈0.13: above a 13 % chance of failing before the stop, stop today.
The acceptable risk comes from the cost sheet.
Three prices (a planned replacement, an unplanned stop, a failure) fix the probability above which waiting costs more than acting; the model's job is only to supply that probability, and a model trained without those prices cannot say when to act.
There are more options than now or later (reduce the load to stretch the life, move the stop forward, stock the spare), and choosing among them in sequence as the indicator evolves is a Markov decision process; the single rule above is its one-step version, the same cost arithmetic as any decision threshold.
Before any of this, the plant needs the basics: an accelerometer per machine, trended limits, and work orders that record the real cause. Predictive maintenance is the cannon of the maintenance ladder, for fleets of identical units, expensive failures, or failure modes with no clear physical signature.