Study On Feature Extraction Of Viral Pneumonia Based On Lung Sounds

The traditional auscultation method has attracted much attention in the diagnosis of lung diseases because of its convenient and simple and non-traumatic characteristics.However,there are also many problems with traditional manual auscultation methods,such as the subjectivity of human judgments on diagnostic results.Especially in the period of the outbreak of the new coronavirus(COVID-19)pneumonia in 2019,traditional prevention and control methods have exposed many drawbacks,the virus is highly infectious,and the diagnosis results are not timely and accurate,resulting in the traditional auscultation method is difficult to cope with the suddenness and complexity of the epidemic.The digitization and automation of lung-sound diagnostics will be a trend in this field.In the process of automatic diagnosis of lung sound,the selection of feature data determines the accuracy of the algorithm to a large extent,so it is the key to the automatic diagnosis of lung sound signal to extract features and find out effective feature values.This thesis is based on the characteristic extraction of viral pneumonia lung sounds: the main contents are as follows:First,the knowledge related to lung sound and the development status of lung sound feature extraction at home and abroad are systematically introduced.The current commonly used feature extraction methods are analyzed,and the advantages and problems are pointed out.Based on the collected lung sound data,this thesis denoises the lung sound data,finds out each respiratory cycle through repeated listening to the lung sound data,and defines the label for each breathing cycle,uses Matlab to write a program to automatically cut according to the generated label,and finally constructs a lung sound database with a label.Second,the data is preprocessed and the lung sound signal is carried out in the traditional way such as the feature extraction study of short-term parameters,and when the time domain and frequency domain feature extraction methods are analyzed based on the lung sound,in view of the fact that some feature extraction methods are calculated for each frame of the lung sound signal to obtain the characteristic values.However,the length of the respiratory cycle of different patients is often different,so the length of the eigenvalue vector calculated after framing is inconsistent,which affects the collation of the eigenvalue matrix,and this thesis proposes a eigenvalue homogenization algorithm based on random numbers to effectively solve this problem.Third,based on the analysis of time domain information based on lung sound,according to the time domain waveform map and energy characteristics of lung sound,the respiratory gas phase separation of periodic lung sound data is carried out,and a wavelet packet transformation feature extraction method based on the exhalation and inspiratory phases is proposed;based on the in-depth study of the wavelet transform and Hilbert-Huang transformation technology,a joint feature extraction(WT-HHT)method based on the wavelet transform and Hilbert-Huang transformation is proposed.Experimental classification algorithm is used to classify and identify the eigenvalues obtained by the traditional feature extraction method and the feature extraction method proposed in this thesis,and the results show that the classification recognition accuracy of wavelet packet feature extraction based on breathing gas phase is 86.35%,and the recognition accuracy based on wavelet transform and Hilbert-transgender feature extraction is 91.39%.Compared with the recognition accuracy based on traditional feature extraction,there has been a great improvement.This thesis verifies the effectiveness of the feature extraction method proposed in this thesis.