Research And Design Of Flash Memory Life Prediction Method Based On Deep Learning

In recent years,flash memory has gradually become a mainstream storage device due to its advantages of fast read and write speed,non-volatility,and good shock resistance.Flash memory relies on charge stored in storage cells to preserve data.During use,the program/erase(P/E)operation on the flash memory block causes wear on the oxide layer.With the increase of oxidation layer wear,electrons are more likely to leak out in the original electric field,resulting in that the written data cannot be effectively saved,and the number of bit errors rises rapidly,which seriously affects the life of the flash memory.Therefore,we study the application of deep learning in flash memory life prediction,so as to migrate the data stored in the flash memory in time before the flash memory has bad blocks,which is of great significance to the reliability of the flash memory.We first build 3D NAND flash data sets to obtain the number of bit errors after a certain P/E cycle of flash memory in the physical world for studying the application of deep learning in flash memory life prediction.When building the data set,the data may be missing due to the sampling frequency,the physical failure of the flash block or the test platform,resulting in incomplete input and output of the flash life prediction model.We propose a missing value completion method based on support vector regression model for solving this problem.The input of the model is all P/E cycles with missing values and the corresponding number of bit errors with missing values.The model is fitted and calculated to obtain the missing value completion model.The P/E cycles,which is missed,are input into the missing value completion model to obtain the approximate number of missing bit errors.Experiment results show that the accuracy of filling missing value in the missing value completion model is high,and the average relative error of the number of filling bit errors corresponding to different missing proportions is between 0.7% and 4%.We propose a method for predicting the life of flash memory based on deep learning for solving the problem that it is difficult to predict the life of flash memory accurately.The input of the model is the number of bit errors corresponding to the early P/E cycles of the flash block,and the output is the number of bit errors corresponding to the later P/E cycles of the flash block in this method.We take the data from the training set for iterative training and select the parameters with the optimal prediction effect and the corresponding prediction model.Then we input the number of bit errors corresponding to the early P/E cycles of the flash block in the test set into the prediction model,and calculate the error between the predicted value of the number of bit errors corresponding to the later P/E cycles and the real value.Experiment results show that the accuracy of life prediction in the deep learning model is high,and the average relative error of the number of bit errors corresponding to the number of different input P/E cycles is between 8% and 14%.