Embedded systems · 2023
Industrial Fire Analysis Module
A smart detector that recognises fire in 50 milliseconds and almost never raises a false alarm.
Conventional detectors either sound too late or sound for nothing. This module evaluates smoke, heat and particle data with AI running on the device itself, recognising a real fire faster than the blink of an eye.
Live simulation
A fire starts as gas, long before heat or smoke.
Watch 6 panels in a factory electrical room live. Start smouldering or a loose connection in a panel and see how much earlier AICO Sense catches the gas anomaly than a temperature threshold or a classic smoke detector. One panel was just installed — you can also watch the board learn its environment.
How it works: the nano edge AI on the board learns each panel’s normal gas behaviour (daily cycle, load changes) in 48–72 hours. Every reading is compared with that normal: the CUSUM algorithm (S = max(0, S + z − k), threshold h = 8) catches small, slowly building deviations and the spike detector catches sudden jumps. Temperature is the second confirmation layer. Smouldering produces almost no heat at first, so gas changes hours before temperature and smoke. Values in the simulation are modelled to illustrate this behaviour.
Challenge
Conventional detectors’ false alarm rates of over 15% and response times of more than 5 seconds were leaving safety gaps in industrial facilities.
Our solution
Multi-sensor fusion with AI running on the device. Smoke, heat and particle data are processed instantly; false alarms dropped to 0.01% and response time to under 50 ms.
The story
How we built it
Traditional fire detectors were creating serious safety gaps in industrial environments due to high false alarm rates and slow response times.
In this project, we developed a custom TensorFlow Lite model running on an STM32 microcontroller. The system processes smoke density, temperature change and particle data in real time and makes a decision in under 50 ms.
Communicating wirelessly over LoRaWAN at ranges beyond 2 km, the system meets ATEX Zone 1 requirements and can be used safely in explosive atmospheres. A 4-layer PCB design ensures EMC compliance.
Technologies used
- STM32H7
- Altium Designer
- TensorFlow Lite
- Edge AI
- C/C++
- LoRaWAN
- ATEX
Technical specifications
| Processor | STM32H743 480 MHz |
|---|---|
| Memory | 1 MB Flash, 1 MB SRAM |
| AI model | TFLite, 128 KB |
| Communication | LoRaWAN Class A |
| Battery | 3.6 V 19 Ah Li-SOCl₂ |
| Protection | IP67, ATEX Zone 1 |
| Operating temperature | -40 °C ~ +85 °C |
| PCB | 4 layers, EMC compliant |
AICO Sense
Catches the very first gases of smouldering, before heat builds up or smoke appears.
View solutionPlanning a similar safety project?
If you have a similar need, let’s bring our experience to your project.