Ray Richardson, Chief Technology Officer at Simularity @ MLconf SEA
summary
Technique converts time-series windows to SAX words, running on tiny devices including a 2K RAM prototype, and computes anomaly scores with simple metrics (p value, log-likelihood ratio, or a cheap rate ratio). Applied to hard-drive telemetry using 53 sensors on ~400 drives to learn event signatures, predict failures and estimate time-to-failure; implementation uses PAA preprocessing and optionally self-organizing maps for magnitude-sensitive features. Open-source SAX code and a correlation engine are available and the deployed system successfully identifies actionable anomalies.
AI-written summary · about
chapters (8)
- 0:00Start
- 4:35Finite Symbol Space
- 9:39The Symbolic Aggregate Approximation
- 14:51Computing the Anomaly
- 17:43A P-Value
- 18:42Log-Likelihood Ratio
- 19:44Rate Ratio
- 23:02Event Signature
related talks
52mDevil in the Haystack
31mFinding the Needle in the Haystack: Predicting Storage Device Failures in...
43mBeyond Logs: Real-Time Anomaly Detection with Machine Learning
39mDeveloping Spidey Senses
29mAnomaly Detection
35mObservability’s Sixth Sense: Detecting Anomalies in Metrics
10mPaco Nathan (Liber 118) Interview - OSCON 2014
47m Natural Language Processing & Anomaly detection in Sys call logs