Soft Sensor vs. Hard Sensor: When to Use Which

The question comes up on almost every project: “Should we install a new analyzer, or can we estimate the measurement with a model?”

It sounds like an instrumentation question. It is actually a control strategy question — and the answer depends on factors most engineers don’t consider until they’ve already spent the capital budget.

---

What is a soft sensor?

A soft sensor (also called a virtual sensor or inferential measurement) estimates a process variable from other available measurements using a mathematical model. It has no physical counterpart — no probe in the pipe, no sample to the lab.

The model can be:

• First-principles (mass/energy balance, reaction kinetics)
• Data-driven (regression, neural networks, Gaussian processes)
• Hybrid (physics structure with data-calibrated parameters)

The model runs continuously in the DCS or historian, typically at the scan rate of the existing instrumentation.

---

The comparison that matters

Property	Hard sensor (analyzer)	Soft sensor
Capital cost	€20k–€500k	€5k–€50k (development)
Maintenance cost	€5k–€30k/yr	Low (model recalibration)
Measurement lag	2–60 min (lab); 30 sec–5 min (online)	Near-zero (scan rate)
Availability	85–95% (maintenance, calibration)	99%+ (software)
Measurement noise	Physical noise + sampling	Model uncertainty
Works without model	Yes	No
Covers multiple outputs	One per instrument	Yes (one model, many outputs)
Fails safely	Usually (alarm on failure)	Needs monitoring

---

When hard sensors win

Use a physical analyzer when:

1. The variable is safety-critical. Gas detection, pressure relief, temperature limits near equipment damage thresholds — these need independent physical confirmation. A soft sensor can support the safety system; it should not replace it.

2. The process is too poorly understood for a model. If you cannot write down a mass balance with known parameters, a first-principles soft sensor is not feasible. Data-driven models require sufficient historical variation in the target variable — if your process rarely changes, training data is sparse.

3. Regulatory compliance requires a certified measurement. Environmental reporting, food-grade batch certificates, pharmaceutical release testing — these typically specify the measurement method. A soft sensor cannot substitute for a certified instrument in these contexts.

4. The measurement lag of the soft sensor would be comparable to a fast online analyzer. For variables that change in seconds (combustion O₂, pH in fast reactors), the model update rate and the dynamics of the predictor variables determine whether a soft sensor can match the response time of a dedicated sensor.

---

When soft sensors win

Use a soft sensor when:

1. Measurement lag is the primary constraint. Lab analyzers with 30–60 minute delays make feedback control impossible for anything faster than an hourly process. A soft sensor that estimates quality at process scan rate enables real-time closed-loop control.

2. The measurement point is physically inaccessible. Temperatures inside a sealed reactor, concentrations mid-column in a distillation tower, internal states of a battery — these may have no feasible measurement point at all.

3. You need redundant coverage. A soft sensor built on a different set of inputs than the primary analyzer provides independent failure detection. If both agree, you have confidence. If they diverge, you have an early warning.

4. You want to extend control to variables you cannot instrument. Product viscosity, particle size distribution, polymer molecular weight — these are often too expensive or impractical to measure continuously. A soft sensor makes them controllable in real time.

5. Capex is constrained but you need better process visibility immediately. A well-designed soft sensor can often be built and validated in 4–8 weeks, faster and cheaper than specifying, procuring, and installing a new analyzer.

---

The hybrid approach most plants miss

The most powerful configuration is not either/or — it is the soft sensor running in parallel with the hard sensor.

The setup:

• Hard sensor provides ground truth at its measurement frequency (every 30 min, or once per batch)
• Soft sensor runs continuously between lab measurements
• The hard sensor periodically recalibrates the soft sensor, correcting for model drift

This is called a moving horizon estimator or simply an inferential + correction architecture. It gives you continuous measurement at effectively zero lag, with the accuracy of a physical analyzer.

The cement industry has used this architecture for clinker quality estimation since the 1990s. The chemical industry uses it for column product quality. The steel industry uses it for continuous property tracking in rolling mills.

---

Decision framework in three questions

1. Is the variable safety-critical or regulatory? → Hard sensor required (soft sensor may supplement)

2. Can you write a credible model (first-principles or data-driven)? → If yes, a soft sensor is technically feasible

3. Is measurement lag or capital cost your primary constraint? → Lag constraint → soft sensor; Capital constraint → soft sensor; Neither → hard sensor likely simpler

If you answer “yes” to questions 2 and 3, the hybrid architecture (soft sensor + periodic hard sensor recalibration) usually delivers the best result at the lowest total cost of ownership.

---

How to estimate soft sensor feasibility in 30 minutes

The fastest feasibility check:

1. List the variables you want to estimate (the “outputs”)
2. List the measurements already available in your historian (the “inputs”)
3. Check whether there is sufficient historical variation in the outputs — if the variable never changes, there is nothing to model
4. Check the time constant of the process — if the variable changes faster than your historian scan rate, a soft sensor cannot help

If steps 1–4 look promising, a week of historian data is enough to build a proof-of-concept model and quantify the estimation accuracy before committing to anything.

---

Dr. Rafał Noga specializes in state estimation and soft sensor development for industrial processes. Previous work includes model development for airborne wind energy systems (SkySails Power), automotive control applications (IAV GmbH), and cryogenic superconducting magnet systems (CERN).