The story
Reads sensor data from real beehives and warns beekeepers early when a colony is getting sick.
Turning raw time-series readings from real hives (temperature, humidity, weight, bee traffic) into features and health labels a model can learn from, before benchmarking 5 classical models against a neural network.
How it’s built
Inside SoundHive AI
System architecture
- HOBOS hive sensor data
- → Time-series preprocessing (400,000+ readings)
- HOBOS hive sensor data
- Time-series preprocessing
- → Rolling-window features
- Time-series preprocessing
- Rolling-window features
- → 3-class health labels
- Rolling-window features
- 3-class health labels
- → 5 classical models benchmark (benchmark)
- → TensorFlow neural network (train)
- 3-class health labels
- 5 classical models benchmark
- TensorFlow neural network
- → Streamlit demo on Hugging Face
- Streamlit demo on Hugging Face
Engineering decisions
Decision 01
Features before models
Raw sensor streams (temperature, humidity, weight, bee traffic) became rolling-window features and 3-class health labels first. The model can only learn what the features make visible.
Decision 02
Earn the neural network
Five classical models, including Random Forest and XGBoost, were benchmarked first, so the TensorFlow network had to beat real baselines, not a strawman.
Decision 03
Guard against overfitting
Dropout regularisation on the network, judged on F1 (0.88) as well as accuracy (98.82%), because unhealthy colonies are the rare class that matters most.
Decision 04
Make it usable
A Streamlit demo on Hugging Face Spaces lets a beekeeper try it without installing anything.