Emerging methods
AI in air quality monitoring — where it helps, where it overpromises
Machine learning is a useful set of tools applied to sensor data — not a replacement for the sensors, the calibration regime or the engineer interpreting the result. This page sets out what AI realistically contributes to monitoring today.
Capabilities
What machine learning is good at
Pattern recognition
Recurring profiles across days, weeks and seasons emerge from long time series — useful for separating occupancy effects from ventilation behaviour.
Anomaly detection
Departures from learned baselines flag events that warrant human review. The model does not diagnose; it surfaces.
Sensor fault diagnosis
Drift, stuck values and unphysical correlations between channels are detectable earlier than scheduled calibration alone would catch.
Short-horizon forecasting
Predicting the next few hours of CO2 or PM trajectory supports demand-controlled ventilation and pre-emptive alerts.
Foundations
Models only work on the data they are fed
Every claim a model makes traces back to a sensor reading. Drift, miscalibration or poor placement propagate through the model as confident-looking outputs. The cleanest model architecture cannot rescue compromised inputs.
Training data must represent the conditions the model will operate in. A classroom model trained in autumn will not necessarily generalise to summer ventilation behaviour. Periodic retraining, with refreshed reference data, is part of operating any predictive system.
Explainability matters as much as accuracy. A model that flags an issue without revealing why does not support the human decision that follows. Useful systems pair predictions with the features that drove them.
Method
Measurement, prediction and decision
| Question | Measurement | Prediction |
|---|---|---|
| What is the level now? | Sensor reading, with uncertainty | Not the right question for a model |
| What will the next hour look like? | Not available from sensors alone | Short-horizon forecast |
| Is this sensor behaving correctly? | Calibration check confirms | Anomaly detection surfaces candidates |
| Is this a real event or noise? | Cross-sensor confirmation | Pattern recognition supports the call |
Limits
Where to be cautious
False positives
An overactive anomaly detector creates alert fatigue. Thresholds, deadbands and confirmation logic matter as much as the model.
Hidden assumptions
A model trained on one building's behaviour may misread another. Transfer learning helps; it does not eliminate the need for validation.
Explainability
Predictions without reasons are hard to act on. Choose architectures and tooling that expose contributing features.
Human oversight
Automated systems supplement, not replace, engineers and consultants making judgement calls about people's environments.
Suitable for
Settings where ML adds genuine value
Large estates
Hundreds of sensors generate data volumes where manual review misses patterns ML can surface.
Long-running deployments
Years of data enable robust baselines and meaningful seasonal patterns.
Research and pilots
Sensor performance studies and ventilation strategy trials benefit from analytical tooling beyond dashboards.
FAQ
AI monitoring questions
Discuss an Air Quality Monitoring Project
Scoping predictive analytics, anomaly detection and ML-supported monitoring for UK estates and research deployments.
Speak to a monitoring consultantConnected sensing