Building a safety system for dangerous mines wasn't easy. Here are the main problems we faced and how we fixed them:

• Noisy and Unreliable Sensor Data: Sensors in mines can get dirty or shaky, giving us "noisy" or bad data. This was a big problem because bad data makes the AI give wrong alerts (false positives).
  • How we fixed it: We built a cleaning step (preprocessing) to fix the data before the AI sees it. We also chose a Random Forest model because it is lightweight and handles messy data better than other heavy models.
• Bad Internet Connectivity: Mines are usually in remote places where the internet is weak or cuts out completely. This made it hard to send real-time alerts to the cloud.
  • How we fixed it: We built an Edge Module that works offline. It processes data right at the mine and can sound sirens or log data even if the internet goes down. We also used MQTT to send data because it is very fast and works well even on slow connections.
• Finding Real Map Data (DEMs): Getting exact 3D elevation maps (DEMs) for specific local mines was very difficult because that data is usually private or hard to access.
  -** How we fixed it:** Instead of waiting for private data, we used open-source satellite data from USGS (United States Geological Survey) to train our Random Forest model. This allowed us to build and test the system immediately.