FailureMiner: A Joint Key Decision Mining Scheme for Practical SSD Failure Prediction...
summary
Claims an interpretable pipeline that boosts SSD failure precision by ~38.6% and recall by ~80.5% using production telemetry (70M logs, 85 attributes) from >350,000 drives. Uses random forest plus boundary-preserving downsampling to keep critical healthy samples, ranks per-tree rules with an improved decision-level SHAP impact score, and aggregates them into compact joint key decisions via a contribution score. Extracts a few actionable rule combinations (UECC/DRAM/capacitor patterns), deploys online at Tencent for over a year.
AI-written summary · about
chapters (17)
- 0:00SSD failure prediction intro
- 1:03Challenges in SSD failure prediction
- 2:05Previous studies and limits
- 3:30Data balancing with downsampling
- 4:34Classification boundary samples
- 4:57Random Forest for SSD failure
- 5:59Attribute-level interpretability limits
- 7:05Joint key decision mining proposal
- 7:27Boundary preserving downsampling
- 8:54Noise removal decision selection
- 9:36Extracting joint key decisions
- 10:19Evaluation on real datasets
- 11:43Failure patterns and factors
- 12:25FailureMiner advantages summary
- 12:47High accuracy and production
- 13:09Method interpretability and usability
- 13:29Q&A
related talks
17mMulti-view Feature-based SSD Failure Prediction: What, When, and Why
20mMSFRD: Mutation Similarity based SSD Failure Rating and Diagnosis for Complex...
27mMaking Disk Failure Predictions SMARTer!
31mFinding the Needle in the Haystack: Predicting Storage Device Failures in...
05mGenerating Realistic Wear Distributions for SSDs (Work-in-Progress Report)
43mPredictive Maintenance: From POC to Production With Spark
53mA SMART-er Ceph: Predicting Hard Drive Failure
16mStop Failures Before They Happen: Gradient Boosting