一种基于预训练的固态硬盘RUL预测方法  

A pre-training-based RUL prediction method for solid state drive

作　　者：王小毫 陈雯柏[1] 张波 刘辉翔 王一群 WANG Xiaohao;CHEN Wenbai;ZHANG Bo;LIU Huixiang;WANG Yiqun(School of Automation,Beijing Information Science&Technology University,Beijing 100192,China)

机构地区：[1]北京信息科技大学自动化学院,北京100192

出　　处：《北京信息科技大学学报（自然科学版）》2022年第5期40-45,67,共7页Journal of Beijing Information Science and Technology University

基　　金：北京市教育委员会科学研究计划项目资助(KM202011232023)。

摘　　要：针对固态硬盘(solid state drive,SSD)剩余使用寿命预测中故障样本数据较少的问题,提出一种基于Transformer预训练的固态硬盘剩余使用寿命(remaining useful life,RUL)预测方法。首先使用基于注意力机制的神经网络模型Transformer对存储磁盘中的硬盘驱动器(hard disk drive,HDD)进行预训练,得到一个硬盘驱动器剩余使用寿命初始模型,再用数据量少的固态硬盘数据对该模型进行精调,最后实现对固态硬盘设备的剩余使用寿命进行有效的预测。实验结果表明:建立的固态硬盘剩余使用寿命回归模型在40天以内的寿命预测中可以达到较好的效果。Aiming at the problem of few fault sample data in the remaining useful life prediction of solid state hard disk,a method for predicting the remaining useful life of solid state drive based on the pre-trained Transformer was proposed.Firstly,the neural network model Transformer based on the attention mechanism was used to pre-train the hard disk drive in the storage disk,and an initial model of the remaining useful life of the hard disk drive was obtained.Then,the model was fine-tuned with the solid-state hard disk data with less data.Finally,the remaining useful life of the solid state hard disk device was effectively predicted.The experimental results show that the established regression model of the remaining useful life of solid state drive can achieve good results in the life prediction within 40 days.

关 键 词：固态硬盘 剩余使用寿命 TRANSFORMER 预训练 硬盘驱动器 注意力机制 神经网络 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]