面向时序SMART不平衡数据的硬盘故障预测算法
2025-07-01李国侯雪雪李静陈辉
Hard Disk Failure Prediction Algorithm for Time Series SMART Imbalanced Data
LI Guo',HOU Xuexue¹,LI Jing¹,CHEN Hui² (1.School of Computer Science and Technology,Civil Aviation University of China, Tianjin 30030o,China; 2. Aviation Meteorological Center,Air Trafic Management Bureau of Civil Aviation Administration of China,Beijing 100015,China)
Abstract: In response to the issue of poor fault prediction caused by the scarcity of data center hard disk failure data,a hard disk failure prediction algorithm that could solve imbalance problems through data augmentation was proposed based on the temporal features of self-monitoring analysis and reporting technology(SMART) data information. The algorithm employed long short-term memory networks to improve traditional generative adversarial networks,and sequence segment data containing fault deterioration trend information was generated to address the imbalance problem in the dataset. Meanwhile,to further enhance predictive performance,the prediction model was integrated with temporal attention mechanism and feature attention mechanism, exploring the sensitivity of different SMART features and time steps to the deterioration process of hard disk failures. Additionally, various typical feature selection methods were combined in the feature selection stage to select key features. Experimental validation was conducted on a real hard disk dataset,and the results indicated that the accuracy,recall and F1 values of the proposed algorithm were significantly improved.
Key Words:imbalanced data;data augmentation;hard disk failure prediction;generative adversarial network;attention mechanism
0 引言
作为经济社会数字化转型的存储保障设施,云存储系统通过硬盘(harddiskdrive,HDD)来提供数据存储服务[]。在云存储系统运行过程中发生硬盘故障不可避免,硬盘故障将导致存储数据丢失甚至数据中心的严重瘫痪,造成极大的人力、物力损失[2]。为预防存储数据的丢失,传统的云存储系统采用基于副本的被动容错机制[3-4],但随着存储系统规模的增大,被动容错机制需要创建大量副本,严重消耗存储资源且无法满足数据中心的高可靠性和高服务性要求。因此,主动容错机制成为国内外学者的重点研究内容。
BackBlaze公司公布的硬盘自我检测分析与报告技术(self-monitoring analysis and reporting technol-ogy,SMART)日志数据与硬盘故障之间存在明显相关性,可以通过分析SMART数据实现硬盘故障的主动容错[5-I]。但硬盘在运行周期内大多处于正常状态,故障发生概率较小,SMART数据中健康样本、故障样本高度混合,造成了严重的正负样本数据不平衡问题。同时,在实际存储场景中,硬盘故障是从健康状态到故障状态的逐渐恶化,并非立即发生故障[12]。由于SMART数据仅在故障当天被标记为故障数据,造成故障前一段时间的恶化数据被错误标记为健康数据,加剧了样本数据不平衡问题。为解决数据不平衡问题,研究人员分别基于以下角度提出了有针对性的平衡方案。
基于数据的角度,研究者使用过采样或欠采样技术来平衡正负样本。Tanaka等[13]通过生成对抗网络(generativeadversarialnetwork,GAN)生成模拟数据,并在基准数据集上证明了生成数据的可靠性。……