FINDING BAD BEES IN THE HIVE
summary
Platform processes tens of millions of devices and reduces network usage into hourly rates (168‑dim week) with only 2–20 observations per day per device. Devices get grouped by density clustering (chosen HDBSCAN + GMM), models train from scratch weekly in SageMaker, real‑time inference via Lambda reading Kinesis Data Streams, preprocessing in AWS Glue. Built a device simulator to validate scenarios; system prioritizes low false positives and adapts continuously, and it works at scale in production.
AI-written summary · about
chapters (10)
- 0:00Introduction
- 0:42Detecting IoT anomalies
- 3:52Generic anomaly challenges
- 7:21Cloud-native architecture
- 9:07ML design drivers
- 10:32ML pipeline and clustering
- 13:45Gaussian Mixture Models
- 15:06Validation and testing
- 17:34Production ML lessons
- 19:20Summary
related talks
01h 00mBeeIoT - Enter the World of Bees
24mWhat's the Buzz with Machine Learning
36mDetecting Honeybee Swarms Using the Integration of OpenCV, Pandas, AI, & PyTorch
41mPeer Beneath the Surface of the Earth
52mDevil in the Haystack
37mBig Data Analytics, Machine Learning, Real Time Processing
13mLoRaWAN-Driven Beehive Monitoring with Students
49mA Real-World Example of Using Network Behaviour to Detect Cyber-Compromise