Sensor fusion

Signals into insight.

Connected-building sensor fusion

Local occupancy prediction turns building sensor data into more responsive heating and ventilation.

YF / 03 Sensor fusion Connected-building sensor fusion — conceptual system illustration
System concept Illustration / 03
Industry
Commercial buildings
Our role
Sensor integration & edge inference
Core tools
Raspberry Pi · ESP32 · MQTT
01 / The challenge

A practical problem.

Fixed heating schedules wasted energy in empty rooms, while basic motion sensors missed changes in occupancy. The system needed to work across buildings with different controls and unreliable connectivity.

Engagement
6 months · Aug 2024–Jan 2025
Project setting
12 office buildings · London
02 / Our approach

Built around the constraints.

  1. Connect the signals.

    Environmental readings, motion sensors, and door contacts are combined at local gateways.

  2. Predict what comes next.

    A lightweight model forecasts room occupancy 15 minutes ahead on Raspberry Pi hardware.

  3. Act locally.

    Zone-level predictions adjust building controls, with reactive operation when confidence is low.

03 / The system

How it comes together.

Components → processing → output

Distributed sensors feed local gateways, where occupancy predictions become zone-level building-control decisions.

  1. Collect

    ESP32 nodes · Environmental sensors

    CO₂, motion, temperature, light, door, and acoustic-level readings capture how rooms are being used.

    OutputTime-stamped room readings

  2. Combine

    Raspberry Pi · MQTT · Filtering

    Gateways smooth noisy signals, handle missing readings, and build rolling statistics alongside calendar features.

    OutputRoom-level occupancy features

  3. Predict

    XGBoost · 15-minute forecast

    A lightweight model predicts occupancy bands. Readings from overlapping sensor zones improve robustness.

    OutputOccupancy forecast & confidence

  4. Adjust

    Building controls · HVAC setpoints

    Forecasts adjust heating and ventilation by zone. Low-confidence predictions fall back to current sensor readings.

    OutputZone setpoints & energy feedback

04 / The outcome

Results from this project.

Occupancy prediction
97%Room-level, with a 15-minute lookahead
Local inference
<5 msOn Raspberry Pi 4
Less HVAC energy
18%Weather-normalised, year-on-year comparison

Before

Fixed HVAC schedules heated and cooled rooms regardless of actual occupancy.

After

Local forecasts adjust heating and ventilation by zone, with reactive control when confidence drops.

In context. Energy use was compared over three months after deployment. Layout changes temporarily reduce prediction accuracy; sensors require periodic recalibration.

Your next project

Working on something similar?

Tell us about your system and where you need support.

Discuss your project