industrial//maintenance//prognostics//health indicator

A health indicator is a measured or computed quantity that changes steadily as a unit degrades and reaches a known value when it can no longer do its job, and it is the variable a prognosis tracks: the vibration RMS of a bearing, the capacity of a battery, the thickness of a brake pad, the efficiency of a pump, the crack length in a structure. The quality of every remaining-life estimate depends on it more than on any algorithm placed on top.

A health indicator is a measured or computed quantity that changes steadily as a unit degrades and reaches a known value when it can no longer do its job, and it is the variable a prognosis tracks: the vibration RMS of a bearing, the capacity of a battery, the thickness of a brake pad, the efficiency of a pump, the crack length in a structure. The quality of every remaining-life estimate depends on it more than on any algorithm placed on top.

A good indicator has three properties. It is monotonic, moving in one direction as damage accumulates instead of rising and falling with the operating point. It is sensitive early, so that it starts to move well before failure and leaves a long P-F interval. And it is consistent across units, so that the same value means roughly the same damage on every pump of the type. Raw sensor readings rarely have all three, because they mix damage with load, speed and temperature; the indicator is usually computed, a vibration band energy normalized by speed, a motor's temperature rise over ambient divided by the square of the current, a pump's power rescaled to a reference speed (physics-based features).

Find the indicator before choosing the model.

A quantity that tracks damage and is cleaned of the operating point makes prognosis a matter of trend and threshold; a modern network fed raw signals that do not carry the damage cannot recover what is not in them.

An indicator can be estimated rather than measured. A battery's remaining capacity is never read directly: an extended Kalman filter infers it from current, voltage and temperature, and an adaptive controller's estimated thrust efficiency is a health indicator of a drone's propulsion (state of charge estimation).

The failure threshold placed on it is more convention than physical law (stop at 11 mm/s), chosen from standards, experience or the point where performance becomes unacceptable; it is one of the sources of uncertainty in the remaining useful life.

Several weak indicators can be fused into one, by a weighted sum, a principal component or a learned model trained to be monotonic; the result is easier to track than any of its parts, at the price of a number no technician recognizes (degradation model for how it is projected).