A soft sensor, or virtual sensor, is software that estimates a quantity that is hard or costly to measure directly by calculating it from other measurements that are available in real time.
Many important variables are measured rarely or slowly: product purity from a lab sample every few hours, melt index from an offline test, moisture from a delayed analyzer. A soft sensor predicts that variable continuously from signals the control system already has, such as temperatures, pressures, flows and power draw.
There are two broad families. First-principles soft sensors use physical models, mass and energy balances or state observers such as the Kalman filter. Data-driven soft sensors are trained on historical data, using methods from linear regression and partial least squares to neural networks. The training target is usually the lab or analyzer value, aligned in time with the process data.
Soft sensors degrade. Catalysts age, feedstocks change, instruments drift, and the relationship the model learned stops holding. Good practice includes comparing predictions with each new lab result, bias correction, scheduled retraining and a clear rule for when operators should ignore the estimate. A soft sensor used inside a control loop, sometimes called inferential control, needs more validation than one shown only on a trend screen. Poor time alignment between lab samples and process data is a common source of error.
Questions about Soft sensor
What do you need before a data-driven soft sensor works?
Historical process data and the lab or analyzer values the model should predict, aligned in time. Poor time alignment between lab samples and process data is a common source of error.
Why do soft sensors lose accuracy over time?
The process changes. Catalysts age, feedstocks change and instruments drift, so the relationship the model learned stops holding. Good practice is to compare predictions with each new lab result, correct bias and retrain on a schedule.
Soft sensor in real plants
- Parkland and Seeq: Parkland's Burnaby refinery uses Seeq for daily monitoring and incident investigation of its main air blower, sitewide inferential modeling, boilers, wet gas compressor and flare stack.
Related terms
Advanced process control (APC)
Advanced process control (APC) is a set of control techniques layered above basic regulatory control, most commonly model predictive control, that coordinate multiple variables at once to hold a process closer to its economic and operating limits.
Process historian
A process historian is a time-series database that collects, compresses and stores data from plant control systems, such as sensor readings, setpoints and alarms, so it can be trended, analyzed and retrieved later.
Digital twin
A digital twin is a virtual model of a physical asset, process, or system that is kept in sync with its real counterpart through live data.
Anomaly detection
Anomaly detection is the identification of observations or patterns that deviate significantly from normal behavior in data.
Sources