迁移学习在3D NAND闪存温度敏感可靠性预测中的应用OA

A temperature-adaptive reliability prediction model was proposed in this study.Based on transfer learning,relatively easy-to-obtain normal-temperature or constant-temperature test data were utilized by the model to construct a preliminary reliability prediction model,with the core characteristics and failure patterns of flash memory reliability captured.On this basis,through transfer learning techniques,migration operations were performed using a small amount of data collected under changing temperature conditions to enhance the model's generalization ability in different temperature environments and strengthen its adaptability to variable-temperature environments.The variable-temperature prediction error is reduced from 6.16× 10-5 to 3.49× 10-5,with a 43.3％improvement in prediction accuracy.It is demonstrated by experiments that stable prediction performance is maintained by the model within a wide temperature range from-40℃ to 85℃,and good adaptability to temperature fluctuations is exhibited by the model.While ensuring the same prediction accuracy,the training overhead required by transfer learning is significantly reduced.