Title：Model training method and device for sleep staging
Country：China
Patent No.：202010963626.8
Legal Status：Under review
Inventor：Milin Zhang, Yiqiao Liao, Xiang Xie, Zhihua Wang
Assignee：Tsinghua University
Address：Tsinghua University,Haidian District Beijing 100084, China
Filing Date：2020-09-14
Issue Date：
Abstract：
The invention discloses a model training method for sleep staging. The method comprises the following steps: constructing a single-channel input competition student model and a multi-channel input teacher model for sleep staging; inputting the multi-channel signals into a teacher model, obtaining teacher model loss according to the output of the teacher model, inputting the single-channel signalsinto the competitive student model, and obtaining competitive student model loss according to the output of the competitive student model; determining knowledge migration loss according to the outputof the hidden layer in the teacher model and the output of the hidden layer in the competitive student model; optimizing the teacher model and the competitive student model according to the competitive student model loss, the teacher model loss and the knowledge migration loss until a stop condition is met; and constructing a single-channel input cooperative student model for sleep staging according to the trained teacher model and the trained competitive student model, training the cooperative student model based on a knowledge distillation technology, and finally obtaining the teacher modeland the cooperative student model with improved precision.
Patent Certificate：
PDF/Jpg