A Method of Soft-Sensing Log-Likelihood Ratios Based on Broad Learning System for NAND Flash Memories

Abstract

To accelerate the convergence of the soft-decision decoder and improve the reading performance of the NAND flash memory, a multi-level reading method is usually adopted to sense precise log-likelihood ratios (LLR). However, multi-level readings interfere with the threshold voltage in the cells, increasing the probability of bit errors and accelerating cells wear. To solve this problem, this paper proposed an LLR soft-sensing method based on a broad learning system (BLS) to replace multi-level reading, which improves accuracy of sensing from 97.1% to 97.3% by using the raw bit error rate (RBER) as one of the inputs. The output classification probabilities of the network model are used to calculate the LLR. Compared with the multilayer perceptron model, the adoption of BLS hugely decreases the computation amount of network training from 113 epochs to 5, making on-chip retraining feasible. The proposed BLS-based LLR soft-sensing method will be of great application potential in the intelligent error control of non-volatile memories.