| With the development of data sharing technology,the size of data increases exponentially,and the form of data storage is gradually moving to the cloud.The difficulty of data security management has doubled,resulting in frequent failures of enterprise data centers.These failures can bring immeasurable economic losses.The report [1,2] pointed out that among the various failures in the cloud storage center,the hard disk problem accounted for the highest proportion.Therefore,reliability research through hard disk data can effectively avoid the occurrence of data disasters.The main research in this thesis are as follows:(1)Aiming at the problem that the unbalanced hard disk sample data and SMART feature redundancy lead to poor true positive rate and false positive rate indicators of the prediction model,a hard disk failure prediction model based on GHM-XGBoost is proposed.First,the m RMR feature selection method is used to reduce the dimension of the hard disk SMART attribute;secondly,the GHM sample equalization method is used to calculate the gradients of various samples of the hard disk to balance the influence between the model results;finally,the parameters of the model are adjusted based on the genetic algorithm Optimized processing.The experimental results show that the proposed GHM-XGBoost hard disk failure prediction model is better than other prediction models in terms of TPR and FPR.The TPR is increased to 0.864,and the FPR is reduced to 0.032.(2)Based on the actual operation and maintenance situation,to settle the problem of a single prediction result of hard disk failure in the data center,a hard disk health degree prediction method based on TCBN(Tree Combined Bayesian Network)combined model is designed.Four single classifiers are trained and the output is used as the feature node,and the "SMART-Degree" hard disk health degree classification is proposed.The statistical health degree level is used as the classification node of the network,and the TCBN combination model is formed by a full connection.The experimental demonstrates that compared with the traditional equal division of health,the accuracy of the optimal interval is improved to 75%.In addition,the TPR of the model is higher than that of the single model in terms of failure prediction,and it can better fit the actual remaining service life,which proves that the model can provide more multivariate failure prediction results with higher accuracy and flexibility.This issue starts from the perspective of the hard disk failure prediction model itself and the actual operation and maintenance scenarios.On the one hand,it provides a refined optimization scheme for the two-classification model of hard disk failure,thereby improving the indicators of the prediction model;On the other hand,it provides a variety of hard disk status information,thereby reducing the frequency of data loss events caused by hard disk failures.The two studies complement each other and provide a variety of hard disk failure prediction solutions for data centers of different sizes,mechanisms,and needs. |
