Virtual sensors are made for measuring values, which can only be measured extremely hard, very expensively or even impossibly by conventional physical sensors. For this purpose mathematical models and software functions are used to calculate the desired values from more simple and cheap available sensor signals.
Machine Learning procedures are ideally suitable for the development and calibration of such sensors, because one only can train the desired functionality in an easy and efficient way. The according procedures also allow the nearly automated identification of the best base sensors for the virtual measurements.
Typical examples of virtual sensors are the evaluation of collisions probabilities and danger levels in automotive safety systems as well as the estimation of traffic capacity reserves dependent on the current traffic situation.