robotics//sensor//virtual sensor

A virtual sensor is a piece of software that estimates a quantity no instrument measures, by combining a model with the measurements that are available, and it is how a plant knows the winding temperature of a motor, a refinery the composition of a stream between laboratory samples, or a drone the wind that pushes it. In the process industry it is called a **soft sensor**. Its output is used like a reading: it drives a protection, feeds a controller, fills a dashboard.

A virtual sensor is a piece of software that estimates a quantity no instrument measures, by combining a model with the measurements that are available, and it is how a plant knows the winding temperature of a motor, a refinery the composition of a stream between laboratory samples, or a drone the wind that pushes it. In the process industry it is called a soft sensor. Its output is used like a reading: it drives a protection, feeds a controller, fills a dashboard.

Nobody measures the temperature of a motor's winding, but the current (which heats it), the temperature of the casing and a thermal model of how heat flows between them reconstruct it well enough to protect the insulation. No sensor on a drone measures the wind, but to hold position in a steady wind the drone must lean into it, and that sustained tilt, through a model of drag, gives the wind speed and direction. In both cases the information was already in the measurements; the model is what reads it out.

Whether it can work at all is a question of observability, asked before choosing an algorithm: if the model does not tie the unmeasured quantity to what is measured, no estimator recovers it, and the next step is a real sensor. The machinery is that of unmeasured state estimation, often a Kalman filter or a Luenberger observer built on a physical model.

When the physics is unclear, the model is learned from data: a regression or a recurrent network (RNN) trained on periods when the quantity was measured, during a test campaign or from laboratory samples. That buys a virtual sensor where no equation exists, at the price of trusting it only inside the conditions it was trained on.

It fails quietly. Outside its operating envelope (a new product, a worn motor, a fouled heat exchanger) it keeps producing smooth, plausible numbers that are wrong, so a good soft sensor is checked periodically against the real quantity, with the laboratory sample or the occasional direct measurement, and recalibrated when they part (data drift).