Exploit both SMART Attributes and NAND Flash Wear Characteristics to Effectively...
summary
Claims device-level NAND wear histograms plus SMART logs improve SSD failure forecasts, using manufacturer-extracted wear tests and 20,000+ self-test logs to build an Aging-Aware Pseudo Twin Network that combines an MLP for SMART series, an RSTM for wear histograms, and a Random Forest adaptive discriminator. Evaluated on 3D TLC drives from the Alibaba T dataset, it adds SMART+NAND wear histograms and online learning, boosting F1 by ~51% and TPR by ~40% and reducing false positives and MTTDL cost significantly.
AI-written summary · about
chapters (9)
- 0:00SSD failure prediction intro
- 1:44SMART attributes limitations
- 3:18NAND flash wear study
- 6:15SSD failure impactors
- 8:04Extracting NAND flash wear data
- 11:25Aging-aware Suda Twing Network design
- 15:46Prediction performance evaluation
- 18:20Online learning for rare SSDs
- 20:28Conclusion and future work
related talks
31mFinding the Needle in the Haystack: Predicting Storage Device Failures in...
20mHDDse: Enabling High-Dimensional Disk State Embedding for Generic Failure Detection
27mGeneral Keynote Multi-tiered digital twins with what-if analysis...
43mPredictive Maintenance: From POC to Production With Spark
22mImproving Reliability by Data-Driven Engineering & Culture Changes
16mPolarStore: High-Performance Data Compression for Large-Scale Cloud-Native Databases
24mManipulating Data in an Artificial Intelligence World
46mML Use Cases at Think Big Analytics