Numerical Simulation Of Acoustics In Bronchi And Lung Sound Classification

In recent years,due to the high incidence and long treatment period,respiratory diseases have become a serious medical problem.It is of great significance to use intelligent information technology to realize automatic classification of lung sound to assist in the diagnosis of lung diseases.Lung sound contains abundant information about the lung physiological and pathological conditions.However,most lung sound classification systems still adopt the parameters commonly used in the field of speech processing,and the characteristics of lung sound signal have not been considered properly.Based on the mechanism of lung sound,this thesis analyses the characteristics of air flow and sound pressure spectrum curve in the lung bronchi,and realizes the endto-end lung sound classification systems based on deep neural network.The systems’performance is verified on the self-built lung sound database and public database.The main work of this thesis is shown as follows.(1)Based on the theory of fluid mechanics the fluid structure coupling model of the bronchi was established to simulate the gas flow in the bronchi.Then the acoustic simulation was established by using aeroacoustics,and the sound pressure level spectrum curve in the bronchi was obtained,which provides ideas for the adaptive processing of acoustic characteristics in the subsequent automatic classification system.(2)An end-to-end framework of lung sound classification based on deep neural architecture is designed.Different network architectures(CNN,TDNN,ResNet,etc.)are adopted as feature encoders to optimize the classification performance of the endto-end systems.(3)Combined with the characteristics of bronchi acoustic simulation,the attention mechanism of feature channel is designed,and the feature encoder is optimized to enhance the channel information in the lung sound data which is conducive to distinguish different categories,and to weaken the channel information which has less contribution to classification.(4)Combined with the spectrum characteristics of endobronchial sound pressure level,the attention module of acoustic frequency band is designed to adaptive weight the corresponding acoustic features,which contributes to the performance improvements of lung sound classification.