Active failure prediction of hard disk drives (HDDs) is critical to prevent data loss and spare parts replacement decisions. Existing methods for failure predictions of HDDs always used a binary classifier to distinguish the healthy or failed HDDs and cannot address the problem of variable degradation states. In this article, an adaptive error tracking method is proposed for the HDD failure prediction. This method regards the extracted degradation feature as time serials and uses a state filter to estimate the real-time HDD's health status. Then, the HDD failure online prediction is achieved according to the alarm threshold determined by the adaptive error tracking. The degradation of an HDD is described by a first-order Markov hybrid jump degradation model, and the advanced Rao–Blackwellized particle filter algorithm, together with the expectation-maximization (EM) algorithm, is derived to estimate the model parameters adaptively. Finally, to verify the effectiveness of the proposed method, an accelerated degradation test (ADT) based on the vibration was carried out. And the data from ADT and real data center show that the proposed method performs much better than the previous methods, such as Kalman filter, SVM, MD, and recurrent neural network (RNN) based methods, with respect to failure prediction and the alarm distance, which helps to backup data and optimize maintenance decision costs for users.

中文翻译：

---

基于自适应Rao-Blackwellized粒子滤波误差跟踪方法的硬盘故障预测

硬盘驱动器（HDD）的活动故障预测对于防止数据丢失和备件更换决策至关重要。现有的用于HDD的故障预测的方法总是使用二进制分类器来区分健康或故障的HDD，并且不能解决变量退化状态的问题。本文提出了一种自适应误差跟踪方法，用于硬盘故障预测。此方法将提取的降级特征视为时间序列，并使用状态过滤器来估算实时HDD的运行状况。然后，根据自适应误差跟踪确定的告警阈值，实现HDD故障在线预测。HDD的降级由一阶Markov混合跳跃降级模型和高级Rao-Blackwellized粒子滤波算法来描述，推导了期望最大化算法，以自适应地估计模型参数。最后，为验证所提方法的有效性，进行了基于振动的加速降解测试（ADT）。从ADT和真实数据中心获得的数据表明，该方法在故障预测和警报距离方面，比Kalman滤波，SVM，MD和基于递归神经网络（RNN）的方法具有更好的性能。 ，这有助于备份数据并优化用户的维护决策成本。