基于AE-LSTM的多目标硬盘故障预测方法  

Multi-objective HDD Failure Prediction Method Based on AE-LSTM

作　　者：王东清 张炳会 彭继阳 艾山彬 王兵 姚藩益 芦飞 张凯 WANG Dongqing;ZHANG Binghui;PENG Jiyang;AI Shanbin;WANG Bing;YAO Fanyi;LU Fei;ZHANG Kai(Inspur Electronic Information Industry Co.,Ltd.,Beijing 100085,China)

机构地区：[1]浪潮电子信息产业股份有限公司,北京100085

出　　处：《计算机测量与控制》2024年第5期66-71,79,共7页Computer Measurement &Control

基　　金：山东省基金项目(ZR2019LZH006)。

摘　　要：硬盘故障预测是在故障发生前发出预警,避免数据丢失或服务中断,提高数据中心的可靠性和安全性;然而,大多数故障预测模型将硬盘故障问题转化为二分类任务,忽略了硬盘故障是渐变过程的,并且缺乏故障诊断功能;因此,提出了一种基于AE-LSTM的硬盘故障预测框架,实现多目标任务:硬盘健康状态分级、硬盘剩余使用寿命预测、硬盘故障诊断;采用回归决策树模型智能化对硬盘健康状态进行标记,并通过AE-LSTM模型提取鲁棒的隐藏变量,构建剩余使用寿命预测模型和硬盘健康状态分级模块,根据AE模块的输入输出差异进行硬盘故障诊断;在Backblaze公开数据集上,对比了RF、LSTM和AE-LSTM三种算法,实验结果证实了AE-LSTM算法在多目标硬盘故障预测中的有效性和优势。Hard disk drive(HDD)failure prediction is used to avoid data loss or service interruption,which sends a warning before HDD failures occur,it improves the reliability and security of data center.However,most HDD failure prediction models convert HHD failures into binary classification tasks,ignoring the gradual deterioration of HDD and lacking of fault diagnosis function.Therefore,an HDD failure prediction method based on auto encoder and long short term memory(AE-LSTM)is proposed to achieve the HDD multi-objective tasks of health status multi-classification,remaining useful life(RUL)prediction,and fault diagnosis.The regression decision tree model is used to intelligently label the HDD health status.Then,the robust hidden variables are extracted through the AE-LSTM model,the RUL model and HDD health status classification model are built.The HDD fault diagnosis is implemented by computing the difference between the input and output of the AE module.By evaluating the random forest(RF),LSTM and AE-LSTM algorithms on the Backblaze public dataset,the experimental results show that the AE-LSTM algorithm has the effectiveness and advantages in multi-objective HDD failure prediction.

关 键 词：硬盘故障预测 硬盘故障诊断 剩余使用寿命 长短期记忆单元 自编码器 

分 类 号：TP206.3[自动化与计算机技术—检测技术与自动化装置]