Research On Hard Disk Failure Prediction Technology Based On Deep Learning

With the rapid development of emerging fields such as Io T,cloud computing,and cloud storage etc.,the total number of data produced by human society shows an explosive growth,and most of these data are stored in the hard disks of data centers.Since hard disk itself has certain storing characteristics,once it is broken,the data stored in the disk will be lost forever.Although we are able to alleviate the problem of data loss by the redundant backup strategy,we lack of more effective failure risk evaluation method to improve the health status of the hard disk.If we are able to forecast the hard disk failure in data centers,it will be of extreme value to ensure the security and reduce the operation costs of data centers.At present,the study of hard disk failure forecast has already been one of the hot issues of the studies of data center area.In the study of hard disk failure forecast,the most common method is to use relevant machine learning algorithm to establish failure category models.However during the study,the hard disk health assessment strategy used by many scholars is a relatively simpler linear strategy,which is unable to conduct effective evaluation to the health status of hard disks with unstable load.Besides,many scholars haven't paid attention to the timing sequence of the hard disk SMART information when they are conducting studies in hard disk failure forecast.Therefore,the forecast performance of had disk failure forecast models has a huge room for improvement.For the above mentioned problems,the main work and contributions of the thesis are as follows:(1)The thesis presents an innovative strategy based on Maxwell–Boltzmann probability distribution to evaluate the health status of hard disks.Compared with the linear evaluation method,this strategy can fully take advantage of the fluctuations of hard disk SMART data,which are caused by unstable loads.In addition,considering the fact that SMART data distribution is not centralized due to different reasons of hard disk failure,it can more accurately assess the health status of hard disks.(2)In order to make full use of the timing sequence of hard disk SMART data,two different hard disk failure forecast models LSTM_SMART and LSTM_HEALTH based on Long Short-Term Memory(LSTM)are proposed.Both models include three modules: prediction module,health evaluation module and failure determination module.The difference between them is that LSTM_SMART model first predicts future SMART information based on historical SMART information,then evaluates the forecast results through health evaluation module,and finally makes failure determination through failure determination module.LSTM_HEALTH model transforms the historical SMART data into the health of hard disks first,then predicts the future health hard disks directly by using the prediction module,and finally,the fault is judged according to the predicted health..Comparatively speaking,the training speed of LSTM_HEALTH model is faster,while the prediction accuracy of LSTM_SMART model is higher.In order to verify the effectiveness of the health evaluation strategy and failure forecast models proposed in the thesis,two groups of comparative experiments were conducted using open source data sets from Baidu and Backblaze.The experimental results show that the proposed evaluation strategy can more accurately evaluate the health status of the hard disk.In addition,the experimental results show that the proposed hard disk failure forecast model can greatly advance the time of hard disk failure forecast under the condition of guaranteeing the forecast accuracy.