Research On Lifetime Prediction Of 3D NAND Flash Based On Artificial Neural Network

With the rapid development of information technology,storage systems require greater capacity and higher reliability.As a new generation of flash memory technology,3D NAND flash memory improves storage density through three-dimensional stacking,but the number of cycles of program / erase named lifetime is greatly reduced,and the reliability problem caused by the lifetime limitation is becoming more and more serious.The manufacturer adopts the method of random sampling and destructive testing to formulate the nominal value of flash memory chip lifetime,which is quite different from the actual service lifetime.On the one hand,the reliability of flash memory cannot be guaranteed,on the other hand,the storage capacity is not fully utilized,so new methods are urgently needed for testing flash memory lifetime.Artificial neural network can quickly learn the characteristics of complex objects,learning flash memory life characteristics through modeling will help achieve the purpose of predicting the lifetime of flash memory blocks through a small number of non-destructive tests.In this paper,we firstly designed a 3D NAND flash memory test platform and test method,studied the lifetime distribution of flash memory blocks and the actual changes of relevant physical parameters at various stages from the beginning to the end of lifetime.Secondly,we designed the data set preprocessing method,artificial neural network structure and a training method according to the flash memory lifetime characteristics.Finally we learned three models,compared the prediction effects of the three models from the two aspects of learning set and actual prediction results,analyzed the universality of the models,and proposed two methods of denoising and multi-model averaging to optimize the prediction effect of the model.The results of the study show that the lifetime of different blocks of 3D NAND flash memory chips is quite different.The nominal value provided by the manufacturer does not describe the flash memory lifetime very well.There is a waste of storage space in the current use of flash memory.The erasing time,feature programming time,average reading time,and number of original error bits of the flash memory block during the P / E cycle can reflect the lifetime characteristics of the 3D NAND flash memory,and are suitable as inputs for the lifetime prediction model.The normalized value of the mean square error of the three models is less than 0.05,and the actual test prediction error rates are 0.123,0.041 and 0.159,respectively.After comparative analysis,the prediction effect of the single-chip model is better than the mixed model,the model is not universal for other chips,and each chip is suitable for individual modeling.Using the denoising method can improve the prediction effect of the model.The normalized value of mean square error of the single-chip model test set is reduced by 31.58% and 23.07% respectively.Using the multi-model averaging method can improve the stability of the model prediction.The model building method designed in this paper can obtain a more accurate 3D NAND flash memory block remaining lifetime prediction model,which provides new ideas and solutions for the study of flash memory lifetime prediction.