Lung Sound Modeling And Recognition Technology Based On Aeroacoustics

Nowadays,because of a trend toward high incidence and younger age of lung diseases,the pressure of hospital admission have been aggravated.For the daily monitoring needs of common lung diseases,it is significant to carry out the research on the automatic classification and identification of lung sound signals.Based on the pulmonary pronunciation mechanism,this paper aims to deal with the theoretical and technical issues in the automatic recognition system of lung sound,such as pulmonary airflow modeling,effective feature extraction and the improvement of recognition performance.The main contents of this paper are summarized into the following three parts:1.Establish a simulation model of bronchial airflow.Firstly,the pulmonary bronchial structure model was constructed.Then based on the theory of fluid mechanics,the turbulence model was employed to simulate the airflow pattern of normal/abnormal lung bronchial structure under breathing state,and the relationship between experimental phenomena and pathological structure was further analyzed,which was the required flow field model for subsequent lung sound pressure modeling.2.Establish a mapping relationship between the airflow pattern and the lung sound pressure under respiratory conditions.We used the large eddy simulation calculation method and the Ffowcs Williams-Hawkings equation model in the field of aeroacoustics to simulate the sound pressure of lung.Through the analysis of the value and trend of the lung sound pressure curve,the correspondence between the bronchial structure and the spectral distribution characteristics of the lung sound signal was clarified,which was provided as the theoretical basis for the extraction of front-end features of the subsequent lung sound recognition system.3.Based on the deep learning architecture,we proposed an end-to-end lung sound automatic recognition system.The DNN-HMM baseline system and the end-to-end deep recognition system were built respectively.To evaluate the efficiency of the proposed end-to-end lung sound recognition system,we executed the DNN-HMM baseline system and the proposed system on the self-made lung sound database.The results showed that the end-to-end deep recognition system can effectively realize the automatic recognition of lung sound with or without the Adaptive Noise Cancellation(ANC).And the recognition rate is 2.94%higher than the DNN-HMM which under optimal configuration(with ANC)as well as 2.21%(without ANC).At the same time,the end-to-end system with noisy training is found to have better noise immunity.