Beefy ML: Ultra-Low Power Algorithms on Cattle
summary
A deployed system that classifies grazing, ruminating and drinking from IMU data runs on Cortex M4F hardware using Zephyr and Zephyr Power Management, with models trained in PyTorch and implemented via CMSIS DSP. Algorithms process a 5 s accelerometer window and runs in ~20 ms over 5s window, the compact classifier lives in ~3 KB ROM model, is unit‑tested against PyTorch outputs, runs on solar ear tags and collars for months, and shows predictable confusions (drinking vs grazing).
AI-written summary · about
chapters (27)
- 0:00Introduction
- 1:06Why ML on cattle matters
- 3:57IMU data to grazing insights
- 5:21Hardware for cattle monitoring
- 7:03Data collection challenges
- 9:30Field trials and remote monitoring
- 10:35Data processing and machine learning pipeline
- 13:02Low-power algorithm challenges
- 14:45Running on constrained devices
- 17:22Optimizing for embedded
- 20:12Algorithm performance metrics
- 21:34Accuracy with confusion matrices
- 23:05Q&A: Hardware acceleration
- 24:30Q&A: Zephyr updates
- 25:56Q&A: Hardware accelerators
- 27:19Q&A: Sensor inputs beyond accelerometers
- 29:29Q&A: Zephyr SDK and upstream
- 30:11Q&A: Environmental data
- 31:43Q&A: Detectable behaviors
- 33:52Q&A: Device communication
- 34:13Q&A: Impact on cattle
- 34:34Q&A: Ultra-low power in Zephyr
- 35:40Q&A: Training datasets
- 37:04Q&A: Deployment on other animals
- 38:08Q&A: Grazing behavior comparisons
- 39:13Q&A: Practical farmer use
- 40:59Conclusion
related talks
41mBeefy ML: Ultra-Low Power Algorithms on Cattle
25mMachine Learning on Microcontrollers; Tiny, Useful and Fun
52mMuvr
17mAI workflow for large scale deployment of far-edge ML devices
44mMachine Learning on Microcontrollers With Zephyr and Emlearn
19mWhy Machine Learning on Microcontrollers?
26mNext generation machine learning for mobile and embedded platforms
42mPorting an AI Powered Wearable Health Monitor to Zephyr on Open Hardware