APPLICATIONS
Merevlemez-hibák előrejelzése gépi tanulással és tíz évnyi adattal
Researchers with San Jose State University and Vanderbilt University have trained and tested some ML approach on ten years of hard drive failure data. The results are a system that can do a reasonable albeit not stellar job at predicting failure rates for particular SeaGate harddrives.
"The encoder-decoder LSTM posted an RMSE of 0.83 during training and 0.86 during testing over the exhaustive 10 year data while being able to generalize competitively over other drives from the Seagate family," they write.
- Trained an encoder-decoder LSTM on 10 years of S.M.A.R.T data
- Used data from Seagate hard drives deployed in Backblaze datacenters
- Dataset contains date, model, serial number, S.M.A.R.T features, and failure status
Miért fontos?
Papers like this are indicative of a broader trend unfolding all around us - everything which has been digitized is now subject to prediction, and there are increasingly good off-the-shelf prediction models available to make this an ever-easier task. Machine intuition is being intermingled with systems that govern our own reality.