Classification Of Respiratory Diseases Based On Abnormal Audio Signals

In recent years,the situation of prevention and control of new respiratory infectious diseases in China is very serious,various respiratory infectious diseases seriously threaten the life and health safety of the nation and hinder the economic development of the society.Early warning and screening of people suspected of having respiratory infectious diseases is the key to establishing a complete epidemic prevention,control mechanism and winning the battle against various epidemics.Respiratory lesions are distinctive features of early respiratory infections,patients often present with a variety of pathological features such as coughing,shortness of breath,and wheezing due to respiratory tract infections.Using audio signal processing technology to classify and identify people with abnormal breathing patterns is conducive to self-screening and isolation of suspected patients at the beginning of an outbreak,effectively slowing the spread of the epidemic and cutting off the transmission route.This thesis mainly conducts classified experiments around healthy and abnormal people,including COVID-19,cold and fever,the main work is as follows:1)A database of abnormal audio signals was established and based on the Coswara database,the database was constructed by collating the coughing and breathing sounds of people with abnormal health status to provide data support for the subsequent research work.2)An audio silent region cropping method based on spectral entropy method endpoint detection is proposed to automatically crop the silent region of audio signals,the method detect the midpoint of the silent region and the start-end points of the audio signal.This method is applied to the preprocessing of cough audio signal and reduce the proportion of non voice signals by more than 30%,optimize the subsequent feature extraction,at the same time reduce the amount of data in the model.3)The acoustic features of audio signals are extracted and analyzed,such as high order Mel Frequency Cepstral Coefficient(MFCC)and Power spectral density(PSD)of cough and breath.The differences between the two types of features of people are analyzed through experiments.At the same time,the clinical features of patients are selected,for instance muscle soreness,loss of smell,fatigue with weakness,diarrhea,difficulty breathing,sore throat as well as other symptoms,all features with high correlation were selected as the multi-modal input of clinical features.4)A frame feature fusion device based on convolution kernel is constructed.The low dimension audio signal features are constructed by frame fusion of MFCC of cough and breath.The low dimension MFCC,PSD features and clinical features are classified by machine learning classifier.The optimal classification accuracy is more than 83%.Meanwhile,the experiment shows that the accuracy of experimental classification is improved by 6%-8% when the cough sound with silent area was sheared,the necessity of cutting silent area is demonstrated.5)A classification model with multi-modal feature fusion is proposed.Res Net18 was used as the feature extraction network of cough MFCC and breath MFCC,Meanwhile Multi Layer Perceptron(MLP)was used as the feature fusion network of cough PSD,breath PSD and clinical features.This model can control the combination of multi-modal inputs by controlling the weight and splicing of features.The experiment shows that the model achieved good classification effect,the optimal multi-modal input accuracy reached 92.73%,precision and recall reached 90.55% and91.58%,respectively.In this paper,an abnormal audio database is established,an audio silent region cropping technique based on spectral entropy method endpoint detection is proposed to crop the non-speech signals,meanwhile the PSD features,high-order MFCC features of the audio signal are extracted and analyzed,a classification model based on convolutional kernel of frame feature fusion and a multimodal feature fusion classification model to classify and recognize the audio signals of two types of people,and achieved a good classification effect.