Researchers Develop AI System to Predict SSD Failures From Inaccurate Reports

Researchers at Seoul National University of Science and Technology developed a Multiple Instance Learning system that predicts individual SSD failures despite inaccurate group failure reports, the university said in a press release. The research was conducted with Samsung Electronics using SSD data from an Alibaba Cloud data center.

The method groups SSDs reported as failing in the same rack on the same date into failure bags. A temporal convolutional network then analyzes each drive's S.M.A.R.T. data over time to identify likely failures.

With training data containing a 40 percent false failure rate, a conventional model recorded an F1 score of 0.261. The mean pooling version of the new method reached 0.717. It also ranked genuine failures at an average of 1.6, compared with 3.5 for healthy drives incorrectly reported as failed.

The paper, titled "Multiple Instance Learning for SSD Failure Prediction Under Customer Failure Biased Labels," appears in Volume 219 of Computers & Industrial Engineering.