ABSTRACT

The electromechanical design of the HDD (Hard Disk Drive) renders it more susceptible to failures than other components of the computer system. The failure of HDD leads to permanent data loss, which is typically more expensive than HDD itself. The SMART (Self-Monitoring, Analysis and Reporting Technology) system warns the user if any HDD parameter has exceeded the predefined threshold value needed for safe HDD operation. Machine learning methods take advantage of dependence between multiple SMART parameters in order to make failure prediction more precise. In this paper, we present a failure prediction model based on the anomaly detection method involving an adjustable decision boundary. SMART parameters are ranked by the importance and the 13 most significant ones are used as the initial feature set in our model. In the following stage, we optimized the feature set by removing those that have no major contribution to the anomaly detection model, forming the final feature set comprising seven features only. The proposed anomaly detection model achieved 96.11% failure detection rate on average, with 0% false detection rate in ten random tests. The proposed model predicted more than 80% of failures 24 hours before their actual occurrence, which enables timely data backup.

摘要

HDD（硬盘驱动器）的机电设计使其比计算机系统的其他组件更容易发生故障。HDD的故障会导致永久性的数据丢失，这通常比HDD本身更昂贵。如果任何硬盘参数超过了安全硬盘操作所需的预定义阈值，则SMART（自我监控，分析和报告技术）系统会向用户发出警告。机器学习方法利用多个SMART参数之间的依赖关系来使故障预测更加精确。在本文中，我们提出了一种基于异常检测方法的故障预测模型，该方法涉及可调整的决策边界。SMART参数按重要性排序，而13个最重要的参数用作模型中的初始特征集。在接下来的阶段，我们通过删除对异常检测模型没有重大贡献的特征集来优化特征集，从而形成仅包含七个特征的最终特征集。提出的异常检测模型在十次随机测试中平均故障检测率达到96.11％，错误检测率为0％。提出的模型可以在故障发生前24小时预测80％以上的故障，从而可以及时进行数据备份。