industrial//digital twin

A digital twin is a model of one specific asset, kept synchronized with it through its data and used to make decisions about it, and it serves to know a machine's current state and wear, predict its behaviour and try decisions on it without touching the real one. Few terms have suffered more marketing, so the working definition is strict: the twin describes *drone number 127*, a simulation describes *a drone of this model*. A simulator becomes a twin only when an estimator updates its parameters as that particular unit ages.

A digital twin is a model of one specific asset, kept synchronized with it through its data and used to make decisions about it, and it serves to know a machine's current state and wear, predict its behaviour and try decisions on it without touching the real one. Few terms have suffered more marketing, so the working definition is strict: the twin describes drone number 127, a simulation describes a drone of this model. A simulator becomes a twin only when an estimator updates its parameters as that particular unit ages.

A twin has four parts, each a familiar tool: a model of the asset (state-space model, usually a grey-box model); an estimator that synchronizes the model with the sensors and returns the state and parameters of this unit (Kalman filter); a prognosis of where it is heading (remaining useful life); and a decision that goes back to the asset. Without the estimator it is a simulation. Without the decision it is a dashboard. A 3D model spinning on a screen is a render.

An honest twin is the closed loop run slowly.

Measure, estimate, predict and decide, with a cycle of hours or weeks; MPC runs the same loop every 10 ms. The twin is where a physical model and the data that correct it meet with a budget attached (model versus data).

The cleanest industrial example is the battery digital twin in an electric vehicle fleet. The model is an equivalent circuit; an extended Kalman filter tracks the state of charge and the real remaining capacity, as battery management systems already do in production (state of charge estimation); the prognosis is the capacity fade; the decision sets charging limits or the replacement date.

The expensive part is upkeep, the twin recalibration. The asset is repaired, reconfigured and ages; a twin nobody recalibrates ends up as a portrait of the machine on commissioning day, confidently wrong, and its estimator's residuals are the first place that shows.

A full twin pays off when several decisions share the same model and the asset is expensive (aircraft engines, wind turbines, a grid substation). If the only decision is when to change one bearing, a degradation model and a threshold do the job; twins that recalibrate themselves remain mostly research.