Does virtual sensing always require a neural network?

No; physical, statistical and hybrid models are possible.

Does virtual sensing eliminate reference measurements?

Usually reference information remains important for development and checking.

Can an input sensor failure affect several virtual outputs?

Yes, if they depend on that signal.

AI Glossary / Statistics and probability / Intermediate

Virtual sensing

Definition

Virtual sensing estimates a quantity that is difficult, costly or impractical to measure directly by using a model and other available measurements. It is also often called soft sensing or inferential sensing. The model may be based on physical relationships, statistical learning or a combination of both. Its output is an inferred measurement, not a direct reading from a dedicated physical sensor, and it requires validation against suitable reference information.

What it means

Many useful quantities are unavailable continuously: a laboratory test may be slow, a sensor may be expensive, or the measurement location may be inaccessible. Related signals can provide information about the missing quantity. A virtual sensor translates those signals into an estimate. The estimate can support monitoring or control, but its reliability depends on the relationships remaining valid. A model can produce a number even when inputs are faulty or operating conditions differ from those used for development, so availability of an output is not evidence of accuracy.

How it works

Define the target quantity, required timing and acceptable error. Collect synchronized input signals and trustworthy reference measurements, accounting for delays and measurement uncertainty. Fit or specify a model, then validate it on data that represent intended operating conditions without leakage. Deploy preprocessing, estimation and quality checks together. Monitor input health, drift and reference comparisons where available, and define what happens when the system is outside its validated range. Dynamic processes may require lagged signals or state-estimation models rather than a simple instantaneous regression.

Three practical examples

Example 1: A process model estimates a product-quality variable between laboratory measurements. Example 2: A battery-management system estimates state of charge from electrical and temperature information. Example 3: A machinery model estimates an inaccessible internal condition using external measurements and physical relationships. These examples do not imply that a virtual sensor can replace every physical measurement. Reference sensing, calibration and operational safeguards remain important, particularly when errors affect consequential decisions.

Think of it this way

Think of estimating the temperature inside a closed room from nearby readings and a model of heat flow. The estimate may be useful when you cannot place a thermometer inside, but it depends on the model and the surrounding measurements. Opening an unexpected door can break the relationship even while the calculation continues to run smoothly.

Worked example

For a simple illustrative virtual sensor, define estimated quality Q as 2 + 0.5T minus 0.2P, where T and P are measured signals in the units used during fitting. With T = 20 and P = 10, the estimate is 2 + 10 minus 2 = 10 quality units. On four separate validation cases, suppose prediction-minus-reference errors are 1, -1, 2 and -2. Mean error is zero, but mean absolute error is 1.5 and root mean squared error is sqrt(2.5), approximately 1.581. Zero average error therefore does not mean every estimate is accurate. These four cases are only an arithmetic illustration, not enough to establish a deployment error bound, operating envelope or reliable uncertainty interval.

Illustrative case study

In this fictional case study, Cedar Process wants a continuous estimate of a quality property measured only through periodic laboratory checks. Engineers align the laboratory timestamps with the material's earlier process conditions and build a virtual sensor using relevant signals. They validate across operating regimes and retain periodic reference testing. When an input instrument becomes unreliable, the system marks estimates as unavailable or uncertain according to a predefined procedure rather than silently presenting ordinary-looking predictions.

Common misconceptions and mistakes

Common mistakes include calling an inferred value a direct measurement, validating on records that leak future laboratory results, and ignoring process delays. Another mistake is assuming an accurate historical model remains accurate after equipment, material or operating changes. Inputs may themselves drift or fail. Low average error can hide large individual errors, systematic errors in certain regimes or poor tail behavior. Automatic adaptation also needs care because updating from faulty references can make the model worse.

Business use

Businesses can use virtual sensing to improve visibility where direct measurements are sparse or expensive. Define how estimates will be used, how errors affect decisions and what reference checks remain necessary. Track performance by operating regime, not only a pooled average. Keep inferred and directly measured values distinguishable in reports and interfaces. Plan maintenance, recalibration and fallback behavior as part of the sensing system. A useful model is one component of a dependable measurement workflow, not a substitute for understanding the underlying process.

Check your understanding

What quality value does the illustrative formula predict?

Answer: Ten.

What is the mean error across the four validation cases?

Answer: Zero.

What is the mean absolute error?

Answer: 1.5.

Why is zero mean error insufficient evidence of accuracy?

Answer: Positive and negative errors can cancel while individual predictions remain inaccurate.

Related terms

Calibration
Related terms include soft sensor
state estimation
sensor fusion
digital twin
concept drift and measurement uncertainty

Learn next

Learn next how time alignment and process delays affect supervised training data. Compare physics-based observers with data-driven regressions and hybrid methods. Then study validation across operating regimes, uncertainty reporting and drift detection so that the estimate's reliability is monitored alongside its numerical output.

Sources

Sources: Petr Kadlec, Bogdan Gabrys and Sibylle Strandt, Data-driven Soft Sensors in the Process Industry, https://eprints.bournemouth.ac.uk/8498/1/CACE_KadlecGabrysStrandt_2008.pdf; MathWorks, AI with Model-Based Design: Virtual Sensor Modeling, https://www.mathworks.com/content/dam/mathworks/mathworks-dot-com/products/model-based-design/ai-with-mbd-virtual-sensor-modeling.pdf. These primary methodological and implementation sources explain virtual sensing. The regression and validation-error calculations are original and independently checked; the process case is fictional.

Previous: Virtual private server Next: Virtual support vector machine