Health Levels Modeling for SSD Failure Prediction

The increasing adoption of solid-state drives (SSDs) due to their high performance and reliability has made failure prediction crucial for ensuring data integrity and availability. Self-monitoring, Analysis, and Reporting Technology (SMART) is a system for drives that periodically reports various operational parameters that facilitate early detection of potential issues. Although many studies have used SMART attributes for approaching this matter – as a binary problem – we test new ways of predicting SSD failures, considering multiple health levels. In this paper, we first use feature selection for selecting the best SMART attributes as learning features. Then, we test the selected features on several classification models and two different prediction horizons of one month and one year ahead of the failure. The preliminary results effectively validate our approaches to address that problem, mainly in the smaller prediction horizon with non-linear models.