Detecting and eliminating covariate shifts in data for a more robust HDD failure prediction

Prediction of HDD failures has garnered significant attention in research, yet the persistence of covariate shifts in data remains a practical challenge. In this work we introduce a novel approach to training covariate shift detection models without the need for additional real data or artificial shift modeling. Moreover, we propose a comprehensive methodology integrating shift detection, administrator alerts, shift elimination, and HDD failure prediction. Experimental results demonstrate the viability of our real-world implementation.