摘 要：保证大型数据中心服务的可靠性越来越重要，硬盘是大型数据中心中故障率最高的组件。如果能够预测硬盘的故障情况就可以提前对数据进行保护和隔离，避免造成重大损失。然而当前的预测器不能同时有效地提取时间序列的长短期依赖关系，学习样本的有效特征。文章提出了基于注意机制的全卷积注意力模型，该模型能够解决长短期依赖问题，有效识别故障模式。最后在采集的 SMART 日志的数据集中证明了模型的有效性。
关键词：硬盘故障预测；异常检测；注意力机制；全卷积网络
DOI:10.19850/j.cnki.2096-4706.2021.24.013
中图分类号：TP391；TP18 文献标识码：A 文章编号：2096-4706（2021）24-0048-03
Hard Disk Failure Prediction Methods Based on Fcn-Attention
ZHANG Jiahui
(Northeastern University, Shenyang 110819, China)
Abstract: Ensuring the reliability of services in large data centers is becoming more and more important. Hard disks are the component with the highest failure rate in large data centers. If the failure of the hard disk can be predicted, the data can be protected and isolated in advance to avoid major losses. However, current predictors cannot effectively extract the long and short term dependency of the time series at the same time, and learn the effective features of the samples. This paper proposes a fully convolutional attention model based on the attention mechanism, which can solve the problem of long and short term dependency and effectively identify failure modes. Finally, the validity of the model is proved in the data set of the collected SMART log.
Keywords: hard disk failure prediction; anomaly detection; attention mechanism; fully convolutional network
参考文献：
[1] MURRAYJ F，HUGHES G F，KREUTZ-DELGADO K. Machine Learning Methods for Predicting Failures in Hard Drives： a Multiple-instance Application [J].Journal of Machine Learning Research，2005，6（1）：783-816.
[2] ZHANG J，WANG J，HE L，et al. Layerwise PerturbationBased Adversarial Training for Hard Drive Health Degree Prediction [J]. IEEE，2018：1428-1433.
[3] SHEN J，WAN J，LIM S J，et al. Random-forest-based failure prediction for hard disk drives [J].International Journal of Distributed Sensor Networks，2018，14（11），pages 15501477188， November.
[4] LI Q，LI H，ZHANG K. Prediction of HDD Failures by Ensemble Learning [C]//2019 IEEE 10th International Conference on Software Engineering and Service Science （ICSESS）. Beijing： IEEE，2019.
[5] Backblaze. Backblaze Drive Stats for Q3 2021. [DS/OL]. [2021-11-05].https://www.backblaze.com/blog/backblaze-drive-statsfor-q3-2021/.
作者简介：张佳惠（1997.07—），女，汉族，黑龙江伊春人， 硕士研究生在读，研究方向：人工智能、智能运维、数据挖掘。