What a soft sensor does between lab samples

Picture the overhead of a debutanizer, C-101. Product purity has to stay above 98.5 mol %. The lab takes a sample, runs it and reports back, and the next result arrives about 90 minutes later. In between, the operators steer the column on temperatures, pressures and flows, and on experience.

Most of the time that works. Then a feed composition upset arrives between two samples. Purity drifts down, crosses the spec limit, and recovers before the next lab result comes back. On paper the shift looks clean. In the product tank it was not.

A soft sensor exists to fill that gap.

The idea in one paragraph

A soft sensor, sometimes called an inferential or virtual analyzer, is a model that estimates a quality you measure rarely from variables you measure all the time. For the debutanizer, the slow measurement is overhead purity from the lab. The fast measurements are the top tray temperature (TI-104), the column pressure (PI-102) and the reflux flow (FIC-110), all logged by the plant historian every few seconds. The model learns how the fast readings relate to the slow one, then predicts purity every 15 minutes, or every minute if you like.

Why the physics makes it possible

At a fixed pressure, the temperature on a tray near the top tracks the composition of the liquid on it. Heavier components raise the boiling point, so a warmer top tray usually means a less pure overhead. Pressure shifts that relationship, which is why the model needs PI-102 as well as TI-104. Reflux tells the model how hard the column is working to hold the separation.

All of this is familiar to a chemical engineer. Machine learning turns the relationships you already reason about into a number on a screen, updated continuously.

Building one, in outline

1. Line up the data. Pull the historian tags and the lab results for the same period, and match each lab sample to the process readings at the moment it was drawn, allowing for sample delay.
2. Clean it with care. Remove periods when instruments were faulty or the unit was down, and keep the upsets. The upsets are what the model most needs to learn.
3. Start simple. A linear regression on two or three well-chosen inputs is often a strong baseline. Move to gradient boosting or a small neural network only if it earns its place on data the model has not seen.
4. Validate against the lab, over time. Split the data by date, never at random, so the test set looks like the future.
5. Check it like an engineer. Does predicted purity fall when the top tray warms at constant pressure? If the model disagrees with the physics, trust the physics and look again.

In Python the core of the model is short:

from sklearn.ensemble import GradientBoostingRegressor

X = df[["TI-104", "PI-102", "FIC-110"]]
y = df["purity_lab"]

model = GradientBoostingRegressor().fit(X_train, y_train)
predicted = model.predict(X_now)  # mol %

The engineering is everything around those lines: choosing the inputs, aligning the samples and deciding when the prediction can be trusted.

Keeping it honest in service

A soft sensor is a model of the plant as it was when the data was collected. Catalysts age, exchangers foul and operating targets move. Two habits keep it useful:

• Compare every new lab result with the prediction made at the same moment. Plot the difference. A steady drift means the model needs retraining.
• Watch the inputs, not just the output. If FIC-110 goes to manual, or a thermocouple starts to read flat, the relationships the model learned may no longer hold, and the display should say so.

Where it pays

The value is rarely the model itself. It is the hour of off-spec product you now see as it happens, the giveaway you can trim because you trust the estimate, and the lab samples you can spend on the questions that matter.

Course 1 builds a purity model like this one on distillation column data, from first Python to a validated model.