The Classification Of Snoring Sound Based On Acoustic Characteristics

Obstructive sleep apnea-hypopnea syndrome(OSAHS)is a sleep-disordered that seriously affects people’s lives.The diagnosis of this disease in hospitals is mainly through polysomnography(PSG),which is extremely inconvenient and costly.Through a series of studies on snoring,the snoring of patients with OSAHS is classified,and achieved automatic recognition of all categories,and thus assist in the diagnosis of OSAHS disorders.For noise preprocessing of sleep snoring signals,a noise reduction method combining spectral subtraction and Wiener filtering based on subspace projection is proposed in this paper.In noise reduction experiments with noise signals superimposed with white noise of different energies,using this method,the signal-to-noise ratio is about 7 dB higher than the result when using spectral subtraction alone,and is about 3 dB higher than using the Wiener filter alone,and it is also superior to the spectral subtraction and Wiener filtering methods in terms of the mean square error and the coherence with the original signal.In the noise reduction experiment of the real snoring signals with complex background noise,the method proposed in this paper also obtains excellent results,it is proved that it can reduce the noise well and maintain the signal integrity.For automatic recognition and extraction of snoring sounds,this paper uses a combination of dual threshold method and adaptive threshold method to cut out all voiced segments in the sleep sound signal,and proposes to use a convolutional neural network to automatically recognize the snoring sounds to distinguish snoring from non-snoring.Two kinds of convolutional neural networks,which are powerful in the image recognition field,are selected: Alexnet and Googlenet.The frequency spectrum,signal and spectograms of the sound segments are trained respectively.The experimental results show that the recognition effect of two neural networks on the signal map is best in the three kinds of maps.The recognition accuracy rate reached 86.62% and 90.54%,respectively,and Googlenet’s recognition effect for snoring sound was higher than that of Alexnet.For the classification study of OSAHS patients’ snoring,this article merges the two types of events based on the similar behaviors of apnea and hypopnea in the change of respiratory airflow.The snoring is divided into pre-respiratory disorder snoring,snoring during respiratory disorder,post-respiratory disorder snoring and normal snoring,achieve the automatic recognition by using convolutional neural network,support vector machine(SVM)and adaptive boost algorithm(Adaboost),in both spectral characteristics and acoustic characteristics of snoring sound.28 acoustic and image parameters such as signal map,spectrogram,spectrogram second formant frequency,Mel cepstrum coefficient,spectral centroid,energy entropy ratio and octave energy were extracted to analysis,and picked 19 of all the acoustic parameters to recognize the four kinds of snoring sound.The experimental results show that the Googlenet convolutional neural network has limited ability to recognize four kinds of snorings and has the highest recognition rate for snoring spectrogram,reaching 68.46%.SVM and Adaboost have achieved good results in the recognition based on acoustic parameters picked.The recognition accuracy rate reached 92.53% and 90.67%,respectively,which provided a certain basis for the diagnosis of OSAHS through snoring sound.