Research On Human Sleep Monitoring Based On Federated Learning

Sleep is an important periodic physiological activity of the human body,which is closely related to various health problems.At present,polysomnography(PSG)is usually used to provide fine-grained sleep monitoring.However,PSG is complicated to operate and requires professional personnel to accompany,which cannot meet the needs of daily use.Therefore,many researchers have studied more convenient sleep monitoring,but most of the studies do not support monitoring multiple sleep indicators at the same time.Besides,attention should be paid to the security of users’ data to ensure that privacy is not violated in the monitoring process.In this paper,we study human sleep monitoring based on federated learning.The triaxial acceleration sensor in a smart watch is utilized to collect acceleration changes caused by slight movement during human sleep.We can monitor three important sleep indicators(sleep breathing rate,sleep position and sleep duration).And we improve the monitoring ability of respiratory rate and sleep position with neural network model without aggregating all user data through federated learning.We present a sleep respiratory rate monitoring algorithm,which removes the direct component and the high-frequency noise of the data,and obtains the frequency domain data by using fourier transform.A federated neural network model is designed to fuse the data of the three axes to output the user’s respiratory rate.The experimental verification indicates that the average absolute error(MAE)of the proposed sleep respiratory rate monitoring algorithm is within 0.7.Furthermore,a sleep position monitoring algorithm is designed.Based on the extracted characteristics related to sleep position in the data,we design a federated neural network model to classify the user’s sleep position.Experimental results show that the accuracy of the sleep position monitoring algorithm proposed in this paper is more than 90% in the four basic sleep positions.Additionally,the watch used in this paper does not provide the function of sleep duration monitoring,so we implement a sleep duration monitoring algorithm based on the smart watch,which utilizes the data collected by the triaxial accelerometer to classify the user status and accumulate the sleep duration.After experimental verification,the sleep duration monitoring algorithm proposed in this paper is less than 10 minutes in error compared with other commercial-off-the-shelf smart watches.