Research On Classification Of Breathing Patterns Based On Audio Signals

Breathing is an important indicator which reflects people's physiological health.Therefore,respiratory monitoring technology has become a research hotspot in the field of biomedical engineering.Existing respiratory monitoring techniques can be classified into contact type and non-contact type according to the type of sensors.Contact-type respiratory monitoring techniques need to contact the human body,restrict body movement,easily cause discomfort to the patients,and cannot monitor the patients for a long time.Non-contact respiratory monitoring techniques can overcome these limitations.There are many advantages of the audio detection method,such as simple and efficient operation,convenient carrying of the device,which is more suitable for monitoring the respiratory signal of the human body than other non-contact monitoring techniques such as radar techniques and infrared techniques.Based on the audio sensor,this paper extracts the effective feature parameters from the respiratory audio signals,and classifies the respiratory patterns through the classifiers.The main work is as follows:1.The physiological characteristics and corresponding audio time domain signal characteristics of the five respiratory patterns were described.The five respiratory patterns include Normal breathing,Central Sleep Apnea breathing,Cheyne-Stokes breathing,Cheyne-Stokes Variant breathing and Bradypnea breathing.Then,the preprocessing algorithms of the respiratory patterns classification algorithm were introduced,including the framing and windowing algorithms and the noise reduction algorithms.2.A respiratory patterns pre-classification algorithm based on EWT(Experience Wavelet Transform)and approximate entropy was proposed.The envelope of the respiratory audio signal was extracted and decomposed by EWT,then,the decomposed modes were sorted components according to the signal energy and were calculated the approximate entropy value.Finally,the respiratory patterns were classified into normal respiratory pattern and abnormal respiratory patterns by the linear discriminant classifier.3.This paper used the nonlinear dynamics method to analyze the abnormal respiratory patterns on the basis of the analysis of time domain,and the features were extracted which can reflect the difference between different abnormal respiratory patterns.Because ensemble learning methods have significant effects on multivariate classification,this paper used the ensemble learning classifiers to train the extracted features to achieve the classification of abnormal respiratory patterns.4.The experiment was designed and the results were analyzed.According to the two-level classification structure described in the paper,the experimental data was processed,and the classification accuracy rate was 92.2% on the respiratory patterns pre-classification algorithm based on EWT and approximate entropy.In the classification of abnormal respiratory patterns based on ensemble learning,the effects of different types of features and different ensemble learning classifiers on the classification results of abnormal respiratory patterns were analyzed,and the selected best classification model was tested with measured data,the classification accuracy rate was 83.9%.Finally,through comparison,the necessity of firstly classifying normal respiratory pattern and abnormal respiratory patterns for the measured data was verified.