Research On Lung Sound Recognition Method Based On Deep Learning And Transfer Learning

The lung sound is an acoustic signal generated when the human respiratory system exchanges gas with the outside world.It contains a large amount of lung organ information and is an important indicator of the physiological and pathological characteristics of the lungs in clinical practice.Therefore,lung auscultation is prevention.And one of the effective means of diagnosing lung disease.With the continuous advancement of signal processing technology and artificial intelligence technology,the analysis and recognition technology of lung sound signal is developing in the direction of precision and intelligence.How to construct an effective lung sound recognition model has become an important factor affecting the intelligent diagnosis of lung sound.Auxiliary treatment and prevention of lung disease are of great significance.In this paper,the denoising and recognition modeling of lung sounds is deeply studied,and the lung sound recognition method based on deep learning and migration learning is proposed.The main work is as follows:First,mixed noise removal techniques are used to remove noise from the lung sounds.In order to remove the noise in the lung sound,the low-frequency noise is filtered out by the high-pass filter,and the heart-pulse sound is separated by the wavelet threshold method to remove the heart sound component.The experimental results show that the method can remove the low frequency noise and heart sound noise of lung sound better.Secondly,a lung sound recognition method based on migration learning and VGGishBi GRU model is proposed,which combines VGGish network and bidirectional gated loop unit neural network(Bi GRU).For the problem of insufficient lung sound samples,the VGish network is pre-trained using the Audio Set data set,and the parameters are migrated to the VGGish network layer in the target network.In order to extract the temporal features of the lung sounds,the parameters of each layer in the VGGish network are frozen,and the lungs are utilized.The audio data fine-tunes the parameters of the Bi GRU network layer to perform heavy training on the Bi GRU network.The experimental results show that compared with other lung sound recognition methods,this method effectively improves the accuracy of lung sound recognition,especially the recognition accuracy of asthma,and the overall accuracy of the experiment is over 87%.Third,the effects of relevant factors on lung sound recognition results were compared and analyzed.Firstly,the effects of lung sound center sound on the recognition results are compared and analyzed.Experiments show that the wavelet threshold method is used to remove the heart sound and effectively improve the recognition accuracy.Secondly,the influence of different migration source domains on the recognition results is compared.Experiments show that the Audio Set data set is used.As the source domain of migration learning,it has better recognition effect;again,the influence of different weight training methods on recognition results is compared.Experiments show that Bi GRU is better than the full connection layer,LSTM layer and GRU layer.And the effect of the Bi LSTM layer.