Classification Of Lung Sound Signals Based On CNN-Transformer

In recent years,Chronic Obstructive Pulmonary Disease(COPD)has become one of the main causes of death in human beings,which seriously threatens the life safety of human beings.But because the early symptoms of the disease are not obvious,the best time to treat patients is often delayed.Therefore,it is important to diagnose and treat COPD patients at an early stage.In this thesis,deep learning technology is used to classify lung sound signals,which can help in the early prevention and screening of COPD.The main research content of this thesis is as follows(1)To address the problem of noise in lung sound signals,the Improved Variational Mode Decomposition(IVMD)algorithm was used to denoise the lung sound signals.Firstly,a lung sound signal acquisition system was built.Secondly,the IVMD algorithm was used to denoise the lung sound signals of 500 groups of COPD patients from the ICBHI public dataset and 500 groups of healthy people from the laboratory.Finally,according to the correlation coefficients,the Intrinsic Mode Functions(IMF)were selected to reconstructed the lung sound signals and the pure lung sound signals were obtained.(2)To address the problem that a single lung sound signal cannot adequately characterize the lung sound signal,the fusion feature parameters were used to fully extract the features of the lung sound signal.The Linear Predictive Cepstral Coefficient(LPCC)and Mel-Frequency Cepstral Coefficient(MFCC)were extracted according to the time series characteristics and nonlinear non-smooth characteristics of the lung sound signal.Then the above two feature parameters was used to obtain fusion features containing more lung sound signal information,which laid the foundation for the subsequent classification.(3)A CNN-Transformer-based lung sound signal classification model was proposed to address the problems of low accuracy of traditional lung sound classification,high layer count of commonly used deep learning models,poor generalization,and inability to extract long time sequences.The CNN network model was built to classify the lung sound signal by using the feature of convolutional neural network model to automatically extract the features of lung sound signal;then the Softmax function in the original network model was replaced by Support Vector Machines(SVM)to build a CNN-SVM network model to improve the generalization of the network model The CNN-Transformer network model was proposed to improve the accuracy of classification by combining the Transformer with parallel structure and extracting the long time sequence of lung sound signal.The experimental results show that the IVMD algorithm can significantly remove the noise and improve the signal-to-noise ratio of the lung sound signal.The accuracy,sensitivity,precision and F1-score of the proposed CNN-Transformer network model for lung sound signal classification were 95.70%,94.62%,93.87% and 93.88%,respectively,which improved 7.32%,2.49%,5.22% and 1.51% better than the CNN-SVM network model,providing COPD patients with the possibility of early screening.