An approach for measuring health of a system using benchmarks. The approach may incorporate defining benchmarks with sensors providing data of selected conditions of a system, collecting a predetermined amount of most recent data from the sensors for one or more modes of the system, training a model that represents the system, the model having a response variable that is a function of the predetermined amount of the most recent data from the sensors, applying the model trained, to the benchmarks to get a response of the model in a form of the response variable that is the function of the predetermined amount of the most recent data from the sensors, and comparing the response of the model of the most recent time window with a response of the model of a previous time windows to determine health of the system.