In this paper, an automatic sleep state identification and quality evaluation algorithm based on convolutional neural network (CNN) is proposed, aiming at solving the problems of expensive equipment, complex operation, insufficient accuracy and weak generalization ability of the existing algorithms. In this algorithm, EEG, EOG and EMG signals are processed by using the lightweight multi-branch CNN model through the multi-modal data fusion framework, and the information is integrated by introducing the cross-modal attention mechanism and multi-scale pyramid pooling operation to improve the classification accuracy. At the same time, depth separable convolution and global average pooling are used to reduce the model parameters, making it suitable for wearable devices. Combining transfer learning and data enhancement technology, the problem of over-fitting on small sample data sets is effectively alleviated. The experiment uses the extended version of Sleep-EDF and SHHS subset data. The results show that the accuracy of the algorithm in the five-stage classification task is 92.3%, the Macro-F1 score is 89.7%, the Kappa coefficient is 0.89, and the parameter quantity is only 1.7M, and the reasoning time is 78ms, which is significantly better than many comparison methods. Cross-dataset verification further proves its good generalization ability. This study provides an efficient and accurate solution for home sleep monitoring and sleep health management.
Yi Wei (Sun,) studied this question.