Prediction models of bit errors for NAND flash memory using 200 days of measured data

NAND flash memory, with its excellent storage performance, has become a leader in storage media for scientific instruments. With improvements in the manufacturing technique, its reliability is decreasing. Thus, methods for improving the storage reliability of scientific instruments has become a major topic. If a method to effectively evaluate NAND flash error distribution is found, it will provide significant guidance for error correction codes and the wear equalization algorithm to improve the performance of NAND flash memory. It will also be a good method to predict the flash memory life. Based on data measured over 200 days from an experimental platform for the NAND flash, in this study, the polynomial fitting method, an artificial neural network, and a support vector regression were adopted to build a NAND flash bit errors prediction model, and efficacies of the methods are compared. The simulation results show that the different evaluation models have their own advantages in different situations.