Lab results arrive late
A four-hour assay means four hours of production you cannot correct. Off-spec material is found after it is made.
Inferential sensing · Process analytics
Measure what your lab measures. Every minute, not every shift.
Soft Sensor Toolbox turns the historian data you already collect into live estimates of the quality variables you can only check in the lab — composition, viscosity, melt index, octane, moisture and more.
- No new instruments to install
- Models you can read as equations
- Live deployment to the plant floor
Test R²0.94held-out samples
RMSE0.31lab units
EstimateLiveevery 60 s
Lab value vs. soft-sensor estimateLab Estimate
ŷ = 0.62·Tout + 0.41·Pcol − 0.28·Ffeed + 0.17·Treflux + 4.03
Drift watchAlerts when the model and the lab part ways
Plant connectorsOPC UA · MQTT · PI Web API · CSV
Built for process industries
- Oil & gas
- Refining
- Petrochemicals
- Specialty chemicals
- Pharma
- Steel & metals
- Food & beverages
The problem
The variable that matters most is often the one you measure least.
Lab assays take hours. Online analysers are expensive, drift, and need constant maintenance. In between, operators run the plant on temperatures and pressures — and hope the quality follows.
Analysers cost to own
Sampling systems, shelters, calibration and spares make a single online analyser a capital project of its own.
The data is already there
Your historian holds years of process and lab data. A soft sensor learns the relationship — and estimates quality continuously.
Workflow
From a spreadsheet of plant data to a live estimate, in four steps.
- 1
Upload & profile
Drop in a CSV or Excel export from your historian. The toolbox profiles every column, detects the time axis, and flags gaps, constants and outliers.
- 2
Clean & choose variables
Handle missing values and bad ranges, pick the lab variable to predict, and select the process variables that explain it — with guidance at every step.
- 3
Train & validate
Fit PLS, PCR, Ridge, Lasso, OLS or k-NN models on a split you draw yourself. Compare them on parity, residual and importance plots — and read the final model as a plain equation.
- 4
Deploy & monitor
Publish the model as a live endpoint. Plant connectors stream readings in; the monitor tracks estimates against new lab results and warns on drift.
Platform
Everything a soft-sensor project needs. Nothing it doesn’t.
Guided data quality
Missing-share and constant-column checks, outlier review, and one-click fixes — every change versioned, every step reversible.
Proven model families
PLS and PCR for collinear plant data, regularised regression, OLS and k-NN — chosen for interpretability, not black boxes.
Diagnostics that convince
Parity and residual plots, variable importance, learning curves and a held-out test score — the evidence a review board asks for.
An assistant in the loop
Ask in plain language — “drop the columns with more than 20% missing” — and watch the same tools run, explained as they go.
Export the equation
Take the fitted model out as a readable equation and coefficients table, ready for a report, a DCS calculation block, or an audit.
Live deployment
Serve estimates over an authenticated API, connect via OPC UA, MQTT, PI Web API or CSV, and replay history to test before go-live.
On the plant floor
Built to run where the data lives.
Connectors make only outbound HTTPS calls and keep buffering on disk while the network is down — so a soft sensor keeps estimating through an outage, and an on-premises deployment is available for sites that allow no outbound traffic at all.
- OPC UARead tags directly from the control system
- MQTTSubscribe to broker topics at line rate
- PI Web APIPull from your existing historian
- CSV replayProve it on history before it goes live
Industries
Where soft sensors pay for themselves.
Refining & oil and gas
Crude quality, product octane and sulphur, column tray temperatures translated into composition — tighter cut points, less giveaway.
Chemicals & polymers
Melt index, viscosity and conversion estimated between lab samples, so grade transitions finish sooner and off-spec tonnes fall.
Pharma & food
Moisture, potency and blend uniformity tracked continuously for batch release support and process analytical technology (PAT).
Steel, cement & utilities
Emissions, clinker quality and boiler efficiency inferred from standard instrumentation — no new analyser shelters required.
FAQ
Questions, answered.
A soft sensor is a model that estimates a hard-to-measure quality variable from easy-to-measure process variables — temperatures, pressures, flows — trained on historical lab results. It runs continuously, filling the gap between lab assays.
A historian export with a timestamp, the process variables, and the lab variable you want to predict, aligned by sample time. A few hundred matched lab samples is a practical starting point.
PLS, PCR, Ridge, Lasso, ordinary least squares and k-NN — with train/validation/test splits you control, cross-validated diagnostics, and the final model exported as a plain equation.
Through lightweight connectors for OPC UA, MQTT, PI Web API and CSV files. They send readings to an authenticated endpoint, buffer locally during outages, and need only outbound HTTPS.
This landing page collects nothing — it is a static site with no forms, cookies or trackers. The toolbox itself can be deployed on-premises for sites that cannot allow outbound connections.
Get started
Your next lab result is already in your historian.
Open the toolbox, upload a dataset, and see your first soft sensor today.