Research On Disk Health Assessment And Failure Prediction Technology Based On Deep Learning

In the modern storage system,disk failure is one of the most important factor for restricting the reliability of the system.Disk failure will lead to user data loss,system read and write performance degradation.In sever case,it will lead to system failure.Therefore,the way of how to predicting disk failure as early as possible is the key factor of improving system performance.In the classification problems,the machine learning has a prominent performance so using machine learning is the most common method to predict disk failure.Most of the researchers are still using the health assessment strategy with relatively simple linear strategy.This strategy is not stable when predicting disk failure because it is limited for such loading situation.In addition,most researchers do not pay attention to the time sequence of disk SMART data,the traditional machine learning classification model is relatively simple.The prediction results are not good enough for this shallow learning.Therefore,disk health assessment and data time sequence are two issues that need to be addressed.In order to solve the problems occur in the process of disk failure prediction,this paper studies the following aspects according to the existing research situation:(1)This paper presents a health assessment strategy based on the Euclidean distance of SMART data.In order to make full use of the data fluctuations generated in SMART data,this strategy is based on the Euclidean distance of SMART data,then introduces a monotonous function with time,and is trained into the training set as a result.So,when most of the machine learning methods use this strategy,they all have improved the disk failure prediction rate and solved the fluctuation problem which has occurred when disk facing a sudden loading.(2)This paper presents a disk failure prediction model based on LSTM neural network.Firstly,the SMART data is filtered by the feature selection strategy based on the information gain ratio.Secondly,building a neural model based on LSTM.By using this model,we can effectively use the timing of disk SMART data.Finally,after comparing multiple machine learning model,it has proved that the model mentioned in this paper provides a well applicability.In this paper,open source datasets from BAIDU and Backblaze are used.The result of experiment has verified the effectiveness of health assessment strategy and prediction model that provided by this paper.(3)This paper establishes a prototype system of disk failure prediction based on deep learning.This system possess many functions such as collect the data and update the model regularly and send the alert to the operational personnel when the failure occurs.This paper introduces the relationship between the modules of the disk failure prediction system and some problems that are often encountered in this process.In addition,it provides the solution of data collection for follow-up researchers.In conclude,this paper summarizes the current work and looks forward to future work on disk prediction for researchers.