industrial//maintenance//prognostics

Prognostics is the engineering discipline of predicting how much longer one particular unit will keep performing its function, given everything measured about it so far, and it is what turns a monitoring alarm into a date someone can plan a stop around. It differs from reliability in its subject: reliability describes when a population of identical units tends to fail, prognostics describes this pump, this battery, this engine, with its own history and its own load.

Prognostics is the engineering discipline of predicting how much longer one particular unit will keep performing its function, given everything measured about it so far, and it is what turns a monitoring alarm into a date someone can plan a stop around. It differs from reliability in its subject: reliability describes when a population of identical units tends to fail, prognostics describes this pump, this battery, this engine, with its own history and its own load.

Its output is the remaining useful life, and it comes in three grades, ordered by how much they know about the unit. Age-based prognosis reads the remaining life from the population's life distribution conditioned on the unit's age, so every unit of the same age gets the same answer. Degradation-based prognosis follows a health indicator (vibration level, battery capacity, brake pad thickness) with a degradation model and projects it to a failure threshold; it is the main road whenever such an indicator exists. Direct regression learns the map from sensor histories to remaining life from many units run to failure, and it needs exactly the data a well-maintained plant does not have.

Prognosis is estimation looking forward.

The same filters that track a hidden state now (Kalman filter, particle filter) track the degradation and its rate, and then run the model ahead to the threshold many times; the spread of those futures is the answer, so a prognosis always comes with its uncertainty.

The famous public benchmark is NASA's C-MAPSS, simulated turbofan engines run to failure; it is the place to learn the methods, and its distance from real plants is the place where most of them stop.

Maturity is uneven: degradation-based prognosis is in production for aircraft engines, whose makers have monitored their fleets for decades, for batteries and large turbines; learned prognosis from raw sensors is mostly research.

A prognosis is useful only through a decision that uses it, which is predictive maintenance; what it can warn about is bounded by the P-F interval of the indicator.