Virtual Sensors in Industry

Yuri Katser24 May 2025

Problem

Plants must control parameters that are measured rarely, inaccurately or not at all. A laboratory sample may arrive once every few hours, a physical sensor may drift from the laboratory value, and some points cannot hold a sensor. Operators then work blind between samples, and ML teams lack a common way to choose a model per case.

Example from the article: copper content in the froth product of primary flotation is known from laboratory samples about every 40 minutes. A model on setpoints, process parameters and computer-vision features estimates it continuously, so the process can be adjusted every 10 seconds instead of waiting for the next sample.

Solution in brief

Treat a virtual sensor as six different tasks: choose a working sensor, correct a physical one, create one from scratch, distill a complex model, interpolate rare measurements, expand the feature space. Pick the model class by task: simple regression or classifiers for the first two, physical or hybrid models for new sensors, regression for gaps.

Questions this case answers
  • How to replace a rarely measured laboratory variable with a continuous model estimate and choose the model class by the type of sensor task
  • How to detect which of several diverging sensors to trust, or whether a single sensor has failed, when no direct reference measurement exists
  • How to decide which task a virtual (soft) sensor solves at an industrial plant and which model class fits it
  • How to replace a laboratory measurement that arrives every 40 minutes with a continuous model estimate

Tools and methods

  • Virtual (soft) sensor — software estimate of a parameter that cannot be measured directly or in time
  • Classifiers and simple regression — choose between diverging sensors and detect failures
  • Simple regression and ensembles — correct control-system data toward laboratory data
  • Physical and hybrid models — create a sensor where nothing is measured
  • Computer vision — extra features such as bubble size, speed and shine
Continue reading — it's free

Google or a one-time email code. No password.