Research On The Analysis Of Nonlinear Features For Snoring Sound Signals And Its Applications

Snoring is common in adult sleep,and a few people think that snoring is a manifestation of sound sleep.However,recent studies have confirmed that snoring is the enemy of health.Repeated apnea caused by snoring can cause hypoxemia,easily induce hypertension,cardiovascular diseases,and in severe cases,can cause night sleep apnea,which is the main cause of stroke or sudden death at night.Snoring is caused by the vibration of loose or collapsed soft tissues by airflow passing through the narrow upper airway during the snoring process.Therefore,snoring sound signal analysis has become a research hotspot in the field of biomedicine and modern signal processing in recent years,providing a convenient and low-cost method for snoring analysis.At present,the research work on snoring signal analysis mainly has the following problems:first,foreign research work shows that snoring signals have typical weak nonlinear characteristics,and traditional linear analysis methods(time domain,frequency domain,timefrequency domain,etc.)are not suitable for this kind of nonlinear non-stationary signal,the limitation of short-term stability makes it impossible to accurately describe the dynamic characteristics of the snoring signal;secondly,the existing few nonlinear analysis methods of snoring signal mainly verify the nonlinearity of snoring from a qualitative point of view characteristics,did not quantitatively analyze the impact of different upper airway obstructions on the nonlinear characteristics of snoring,which affects the subsequent localization and identification of snoring;in recent years,a few studies have successfully correlated snoring sound with sleep stages,but the accuracy of the final sleep stages is low because of the discontinuous distribution of snoring during sleep.Therefore,how to design the characteristic analysis method for snore signal to improve the performance of snore detection,how to deeply dig the characteristic information of snoring sound with different excitation location,and how to integrate more useful information to improve the effectiveness of sleep staging are the research hotspots at this stage.Focusing on the above research hotspots,this dissertation has carried out research work in the following aspects:1.Research on snoring detection algorithm based on empirical wavelet information entropyThe traditional acoustic and time-frequency characteristics for snore detection can not accurately describe the dynamic characteristics of snoring sound,thus affecting snore detection accuracy.To solve this problem,a new snore characteristic,empirical wavelet information entropy,is proposed.Firstly,empirical wavelet transform method is applied to snore analysis.For spectrum segmentation,an improved empirical wavelet transform method based on morphological filtering is proposed.Secondly,the empirical wavelet decomposition is combined with information entropy as a new characteristic to describe the dynamic changes of snoring sound.Finally,the effectiveness of this feature in snore detection is verified.The simulation and measurement results show that compared with the traditional acoustic features and time-frequency features,the accuracy and robustness of snore detection based on empirical wavelet information entropy are significantly improved.2.Research on the classification of the excitation location of snore sounds based on the nonlinear characteristics of snoringAiming at the shortcomings of the existing nonlinear analysis methods of snoring sound signals,which only from a qualitative point of view,we firstly discuss the methods of extracting nonlinear features of snoring sound signals,including the calculation methods of features based on chaos theory,long-range correlation features,and complexity features;secondly,the difference of the nonlinear characteristics of the snoring signals at different excitation location is established,and the mapping relationship between the nonlinear characteristics of the snoring sound and the excitation location of the snoring is established.Finally,the above characteristics are combined with the traditional features,and the Boruta algorithm is used for feature optimization to put forward the classification algorithm for the excitation location of snore sound.The test results of clinical data show that the algorithm effectively improves the accuracy of identifying the snore sounds' excitation location.3.Research on sleep staging technology based on the nonlinear characteristics of snoringDue to the difference in sleep among different individuals,the distribution of snoring sounds will be discontinuous during the whole night's sleep.The existing sleep staging methods based on snoring signals are unable to estimate the sleep stage of the missing parts of snoring,which limits the practicability of the current methods.Moreover,for the nonlinear and nonstationary snoring sound signal,it is incomplete to use time domain and frequency domain characteristics to characterize it,which leads to poor sleep staging results,especially for fourstate sleep staging(Wake/REM/Light/Deep Sleep).Therefore,we propose a sleep staging technique based on the multi-feature fusion of snoring/breathing sound.Firstly,the subjects' night-time sound signals were collected to separate the breathing sound and the snoring sound,and the characteristic information of the breathing sound and the snoring sound were extracted specifically.Secondly,to effectively combine the time-frequency characteristics of breathing sound,the time-frequency characteristics of snoring sound and the nonlinear characteristics,and realize the accurate mapping of characteristics and sleep stages,an improved canonical correlation analysis algorithm based on Relief F was proposed.Finally,the ensemble learning classification algorithm is used for sleep staging.The validation results of clinical data show that the proposed algorithm can accurately stage four states of sleep and has high robustness.