Soft sensors (virtual sensors)

Some of the most decision-relevant quantities in a plant are the hardest to measure continuously: melt viscosity, concentration, moisture, the final quality grade that only a lab test reveals two hours later. A soft sensor closes the gap with inference: a model trained to estimate the unmeasurable quantity from the dozens of signals you do measure continuously.

From regression to ML

Process industries have built soft sensors for decades from first-principles and linear regression. The ML generation changed two economics: construction cost (models learn the relationship from historian data instead of months of manual modeling) and fidelity (nonlinear methods capture interaction effects — the material-lot-by-temperature-by-speed dependencies that linear models flatten). The result is that predictive quality — scoring in-process product against final quality, continuously — became a standard platform feature rather than an engineering project.

Where they pay

Three recurring wins: replacing lab latency (act on estimated quality now, confirm at the lab later), impossible locations (estimating conditions where no physical sensor survives), and cost avoidance (a model on existing signals versus an analyzer plus its maintenance contract). The discipline that keeps them honest is periodic reconciliation against ground truth — a soft sensor that drifts unmonitored is a confident liar.

FAQ

How accurate are ML soft sensors?

Accurate enough to act on when regularly reconciled against lab or inspection ground truth — that reconciliation loop is a design requirement, not an option. Treat the estimate as a control signal with known error bars, not gospel.

Are soft sensors the same as predictive quality?

Predictive quality is the flagship soft-sensor application: the 'sensor' being synthesized is the final quality result, estimated live from in-process signals so bad runs get caught mid-run.