Research On Error Rate Prediction Method Of Flash Memory Based On Machine Learning

At present,2D MLC(Multi-level Cell)NAND Flash is widely used in various systems from mobile devices to large base stations.However,with the increasing demand for storage capacity,the capacity expansion of 2D NAND Flash has reached the bottleneck,and the 2D NAND Flash has gradually been unable to meet user needs.The large capacity of 3D NAND flash makes it gradually occupy the storage stage.However,the major problems faced by 2D MLC and 3D TLC(Triple-level Cell)NAND Flash are low tolerance of P/E operation.In order to improve its reliability,the dynamic ECC(Error Correcting Code)is used to correct errors.If the accurate error rate is predicted in advance,it can provide guidance for the selection of dynamic ECC.Therefore,this paper uses machine learning algorithm to build prediction model,and predicts the error rate of NAND Flash,which is of great significance to improve its reliability.Most of the research on NAND Flash is based on simulation experiments,and their results are obviously inaccurate.This paper performs actual wear test on the 2D MLC NAND flash and 3D TLC NAND flash chip based on the experimental platform developed by our research team,and analyzes their characteristics according to the experimental data.Based on the experimental data of 2D MLC Flash memory,error rate prediction model of 2D MLC NAND Flash chip is constructed by Elastic Net,BP neural network and random forest algorithm respectively,and the results of the model are compared and analyzed.This paper constructs the 3D TLC NAND Flash model based on short-term and long-term retention data,respectively.The training and testing process of the three kinds of algorithm models are carried out respectively,and the results of the model are analyzed.Finally,the performance of all the algorithm model is compared and evaluated.The testing results of the algorithm model show that the proposed algorithm model predicts the error bits of NAND Flash chip with high accuracy.Moreover,the data used in the model training in this paper are real experimental data,so that the model results are reliable.