Predicting Failures in Hard Drives with LSTM Networks

Several research has been done to propose early failure detection techniques for hard disk drives in order to improve storage systems availability and avoid data loss. Failure prediction in such circumstances would allow for the reduction of downtime costs through anticipated disk replacements. Many of the techniques proposed so far mainly perform incipient failure detection thus not allowing for proper planning of such maintenance tasks. Others perform well only under a limited prediction horizon. In this work, we present a remaining useful life estimation approach for hard disk drives based on SMART parameters that is capable of predicting failures in both long and short term intervals by leveraging the capabilities of LSTM networks.