Making Disk Failure Predictions SMARTer!
summary
Claims 0.95 F-measure and 0.95 MCC for a 10-day horizon, trained on 380,000 drives across 64 sites with SMART, disk-level performance and four-level location markers. Uses J‑index feature selection, compares Bayes, Random Forest, GBDT, decision tree, LSTM and CNN‑LSTM, and reports performance metrics plus location reduce false negatives and improve coverage; without performance metrics tree models match deep nets; cross-site portability improves when training on many sites; dataset and code are public.
AI-written summary · about
chapters (11)
- 0:00Introduction to hdd failure
- 1:37Data collection methodology
- 2:58Data feature analysis
- 5:31Feature selection process
- 8:08Experimental setup
- 10:20Experimental results
- 12:36Model comparison
- 13:53Failure analysis of models
- 16:28Model portability
- 17:20Sensitivity and conclusion
- 18:34Q&A session
related talks
14mFailureMiner: A Joint Key Decision Mining Scheme for Practical SSD Failure Prediction...
17mMulti-view Feature-based SSD Failure Prediction: What, When, and Why
53mA SMART-er Ceph: Predicting Hard Drive Failure
31mFinding the Needle in the Haystack: Predicting Storage Device Failures in...
40mMachine Learning Model Evaluation Metrics
20mMSFRD: Mutation Similarity based SSD Failure Rating and Diagnosis for Complex...
01h 33mUnderstanding and diagnosing your machine-learning models
14mMinder: Faulty Machine Detection for Large-scale Distributed Model Training