An electrode conductivity sensor fails in a small number of predictable ways. It polarises when the current density is too high, it drifts as the cell constant changes with coating or wear, and its temperature correction is wrong when the sample’s composition is not what the compensation curve assumes. Virtual sensor models — soft sensors trained on plant data — are one way to keep a useful conductivity signal in service when the physical measurement is compromised.
Table of Contents
Where Electrode Measurement Struggles
- Polarisation and cell constant drift. At high conductivity the current density across the electrodes rises, and measurement error follows unless the drive frequency and cell geometry are matched to the range. Four-electrode cells address part of this by separating current-carrying and sensing electrodes.
- Coating. Scale, biofilm and oil films change the effective cell constant. The failure mode is seductive because it is gradual: the reading stays plausible while it is wrong.
- Temperature compensation. Conductivity rises roughly 2% per °C for typical waters. The standard compensation curve is a good approximation for natural waters and a poor one for concentrated brines and non-standard industrial streams, which is exactly where the measurement is being asked to do control work.
- Chemical attack. Aggressive streams and CIP cycles degrade electrodes and seals over time.
Line size and installation add to this: an electrode in a dead leg or in a horizontal pipe with gas pockets will read incorrectly no matter how good the electronics are.
What a Virtual Sensor Actually Is
A soft sensor is a model that estimates a variable that is hard to measure, from variables that are easy to measure. In conductivity service that means estimating conductivity (or the correction to the measured conductivity) from correlated process variables — temperature, flow, pH, differential pressure, dosing rates, and the readings of neighbouring instruments — plus the instrument’s own history.
There are three common forms, and the distinctions matter when specifying one:
- First-principles models. Mass and charge balances around a unit operation. They extrapolate sensibly and are transparent, but they need the plant’s chemistry to be known and stable.
- Data-driven models. Regression, PLS or a neural network trained on the historian. They fit local behaviour well and degrade silently when operating conditions move outside the training set.
- Hybrid models. A physical model with data-driven correction on the residuals. This is the form that usually works in industrial practice, because the physical term keeps the model sane at the edges of the data while the data term absorbs the site-specific behaviour.
For conductivity specifically, the useful first-principles component is often a mixing or blending model — for example, estimating the conductivity of a blend of streams from the conductivity and flow of each — with a data-driven term for temperature and composition effects.
Where Soft Sensors Add Value
Fouling detection and calibration scheduling. Instead of calibrating to a fixed calendar, compare the measured value against the model estimate and act on the divergence. The model estimate is not a replacement for a calibration, but a persistent gap between the two is a good trigger: it indicates coating, electrode degradation or a compensation error. In practice this shifts effort from routine calibration to targeted intervention, and it produces a documented reason for each intervention.
Redundancy at critical points. Where conductivity is used in a control or compliance loop, a model running alongside the sensor provides a cross-check. When a tube of instrument air or a blocked sample line takes the sensor out, the model can keep the loop in a safe holding state rather than dropping to manual.
Coverage where an electrode won’t survive. Points where installation is impractical — a heavily fouling stream, a point that cannot be isolated, a temporary measurement — can often be inferred from neighbouring data well enough for trending, even if not for compliance.
What Soft Sensors Are Not
They are not measurements. A virtual value has no metrological traceability, and it cannot satisfy a permit requirement that names a method. They should not be used as the compliance record, and a plant that quietly substitutes modelled values into reported data has created an audit problem rather than solved a measurement problem.
They are also only as good as the correlations they are built on. In an application with stable operation and little parameter variation, there is nothing for the model to learn from: if conductivity never moves except when the sensor drifts, the model will learn the sensor’s drift. Candidate applications tend to be streams where composition varies and several independent measurements respond to it — cooling tower blowdown as cycles of concentration change, RO concentrate at varying recovery, and blending or neutralisation streams.
Implementation Notes
Data prerequisites. A historian with consistent time stamps, documented units, and a record of which readings are temperature compensated. Models built on data whose provenance is unclear will be distrusted the first time an operator disagrees with them.
Retraining. Process conditions change — new membrane elements, a different antiscalant, a change in production mix — and models need retraining against that. Budget for model maintenance in the same way as instrument maintenance, or accuracy will decay quietly.
Placement. Cloud-hosted models suit offline analysis, reporting and optimisation. Models that feed control or protective logic belong at the edge, where they keep running when the network is down. Many plants split the work along exactly that line.
Validation. Hold back a period of data from training and use it to test. Report the error band, not a single accuracy figure, and re-test after every retraining.
Wrapping up
Virtual sensors are a complement to good instrumentation, not a way to avoid it. The applications where they earn their keep are the ones with variable conditions, several correlated measurements and a real cost associated with losing the signal. Where the underlying sensor is well installed and maintained, a soft sensor mostly becomes a useful inconsistency detector — and that alone often justifies the effort, because the gradual failure modes in conductivity measurement are the ones that do the most damage.
