Research On Flash Memory Channel Estimation And Detection Technology Assisted By Deep Learning

In recent years,with the improvement of semiconductor process technology,the storage density and unit capacity of NAND flash memory have been increased unprecedentedly.NAND flash memory has been widely used in electronic consumer products and has become the mainstream storage medium.However,the channel noise of flash memory becomes stronger with the increase of data storage time and wear times,which greatly reduces the reliability of data stored in flash memory,and the data is facing the loss that is difficult to recover.In the early stage of flash memory,the channel noise is small,the traditional optimal read threshold scheme can play a good role in detection,but solving the approximate optimal read threshold depends heavily on the prior knowledge of flash memory channel.The error correction performance of low density parity check(LDPC)codes based on soft decision is better than that of read threshold detection technology.If we can not make full use of the prior knowledge of flash memory channel and calculate the log likelihood ratio(LLR)value with high accuracy,it will cause the consumption of additional computing resources and the degradation of error correction performance.Therefore,this paper proposes a deep learning aided flash channel estimation and detection technology,which can not only improve the reliability of flash data,but also reduce the dependence on prior knowledge of flash channel.In view of this,this paper first studies the basic structure and programming / erasing principle of NAND flash memory,and further analyzes the channel characteristics by combining with the noise generation mechanism and threshold voltage distribution characteristics of NAND flash memory channel.Then it introduces some deep learning algorithms used in this paper,and studies how to use them for noise estimation and detection of flash memory channel.The specific research work and innovation are summarized as follows:(1)The basic principle and structure of NAND flash memory are deeply discussed;the main noise of flash memory channel is simulated,and the general rule of noise affecting the threshold voltage distribution of flash memory is studied.(2)The deep learning algorithm used in this paper is introduced.The iterative architecture of convolution neural network(CNN)and belief propagation(BP)algorithm is improved,and used for flash memory channel estimation and decoding to achieve decoding,which further improves the decoding performance.(3)According to the characteristics of inter cell interference and persistent noise,this paper proposes a flash memory channel detection scheme based on convolution neural network.The principle is that the convolution layer of convolution neural network can extract the threshold voltage characteristics between flash memory cells.In this scheme,only good data sets are needed to train the detection model which adapts to the channel change and complete the detection of the storage state of the flash memory unit.In order to avoid frequent updating of network parameters,this paper also tests the robustness of the model.The network parameters can be optimized periodically in idle time to match the channel,and the performance will not be lost.Compared with the existing detection models,the proposed detection model can obtain lower detection delay.(4)Aiming at the characteristics of inter unit interference and persistent noise,a method of calculating log likelihood ratio based on depth learning assistant is proposed.This method can calculate LLR value online.The experimental simulation shows that the online LLR proposed in this paper can improve the decoding efficiency of BP algorithm compared with the traditional LLR value.