Hard disk drives failure detection using a dynamic tracking method

Hard disk drives (HDDs) are the core components of data center in IT companies. A breakdown of HDD may cause horrible data loss and great economic loss. Therefore, failure prediction for HDDs is significant to avoid loss and make a data backup plan in advance. Existing prediction methods always focus on a fixed threshold to distinguish whether a HDD is healthy or not, and these methods neglect the problem of multi-stage degradation phenomenon of HDDs. To solve these problems, this paper proposes a dynamic tracking method for HDD failure prediction based on a switchable state stochastic process model. By utilizing Rao-Blackwellized particle filter, the model estimates and parameters are updated by newly available data. To improve model ability, a sensitive health indicator is constructed from SMART attributes based on multiple regression analysis. Then, based on the statistical property of the tracking residuals, the dynamic failure threshold is designed to realize the online prediction of HDD failure. Furthermore, experiments of proposed method are carried out in a real-life data set. The results show the validity of the proposed method.