Cough Recognition Based On Wireless Channel State Information

With the development of society,environmental pollution,living habits,influenza viruses and other factors make the population of respiratory diseases increasingly younger.The death rate caused by respiratory diseases is the third in the urban mortality rate,and even the first in the rural area..Everyone’s breathing characteristics are basically the same,and the universality of breathing is the basis for people to study them.Monitoring respiration can provide valuable data for health monitoring,so that doctors can diagnose the condition more effectively.Cough,as a specific respiratory feature,has high reference significance for the diagnosis of various conditions.Most of the current studies are combined Commonly used endpoint detection algorithms in the field of speech recognition,cough monitoring by analyzing the characteristics of the human cough sound.Sound-based detection methods have low recognition accuracy in complex noise environments,and most indoor environmental noises are ubiquitous.With the widespread deployment of Wi Fi networks,the channel state information(CSI)of Wi Fi devices can record changes in transmission channels caused by different breathing actions,which can be used for cough recognition.Cough detection-oriented CSI data collection and preprocessing technology uses linear fitting methods to eliminate phase errors caused by device clock synchronization and detection delays,and replaces outliers of CSI amplitudes with filters based on median absolute deviations As the median,the one-dimensional CSI amplitude data is decomposed into 5 layers for wavelet denoising processing to remove the noise of signal amplification.The feature extraction technology for cough detection uses CSI amplitude to quantify the sensitivity of subcarriers to human motion.In order to extract obvious features and reduce computational cost,the most sensitive subcarrier phase information is screened for processing,and the phase difference between antennas is extracted The characteristics of variance,range,mean,and interquartile moment are used as the input of the classifier.Due to the limited sample data,the classifier uses a prototype network based on small sample learning.The prototype network requires only a small amount of sample data for each category.It maps the sample data of each category into the embedding space,extracts their mean value as the prototype of the class,and trains to make the distance between the data of this category and the prototype of the class the closest,and Other prototypes are far away.In classification,the distance between the test sample data and each class prototype is made softmax,so as to judge the type of the test sample.The experimental results show that in a complex indoor environment,the system has a recognition rate of 93.62% for human cough,which has a higher accuracy rate than the KNN algorithm,and can be used as a cough monitoring program to provide users with a health reference.