Smoke Detector

At the end of Junior year, my house was burned down. That experience made me realize that most smoke detectors only do one thing: detect smoke. They cannot tell you what is burning, how toxic the fumes are, or how quickly the danger is spreading. That lack of clarity stayed with me and eventually turned into a challenge I wanted to solve.

In my junior year, I decided to confront that fear through engineering. I set out to build a smoke detection system that could analyze and interpret a fire rather than simply reacting to it. I wanted a device that provided real data about the fire’s source, toxicity, and intensity, turning chaos into information people could act on.

I started by designing the hardware. I built a circuit that combined multiple chemical sensors, including MQ-series gas detectors for carbon monoxide, methane, and volatile organic compounds. I added a particulate matter sensor (PM2.5) to measure smoke density and a DHT22 module to record temperature and humidity. These sensors were connected to an Arduino microcontroller, which gathered and processed continuous readings from the environment.

The real complexity came from the software. Most smoke detectors use a simple threshold rule: if smoke exceeds a set value, the alarm goes off. I wanted something adaptive. I used Python and TensorFlow Lite to train a machine learning model on data collected from controlled combustion experiments that involved wood, paper, plastic, and synthetic materials. The model learned to identify the source of smoke based on its chemical and particulate signature. After optimization for edge computing, it could run directly on the microcontroller with minimal delay.