Research On Lung Sound Recognition Method Based On Dual-channel Convolutional Long And Short-term Memory Neural Network

Lung sounds are a physiological signal produced when the body breathes,and when they are abnormal they often indicate a pathology of the lung organs.However,the traditional method of lung sound diagnosis requires a very high level of experience and skill of physicians,and therefore is not easy to implement on a large scale.In this paper,by studying the temporal characteristics of lung sound signals and using neural networks to achieve lung sound pathology diagnosis,we effectively lower the threshold of clinical implementation.To address the drawback that a single model cannot acquire both spatial and temporal information on the lung tone signal,this paper first proposes a single-channel convolutional long and short term memory neural network algorithm(CNNLSTM)to serially acquire features of lung tone data.The LSTM layer in this model can learn the features extracted by the CNN layer,retaining useful information and forgetting invalid information,which overcomes the defects of traditional CNN models that are prone to overfitting in small sample training and the defects of a single LSTM model that cannot effectively mine the information among discontinuous data.To further address the problem of feature map resolution degradation due to CNNLSTM model after CNN extraction,this paper proposes a dual-channel convolutional long short-term memory neural network algorithm(DCCLNN)to obtain features of lung sound data in parallel and generate new features by weighted fusion,which effectively compensates for the deficiency of feature map resolution degradation of CNNLSTM after CNN extraction.The main contributions of the paper are as follows:(1)In this study,Butterworth filter denoising,data normalization,wavelet transform,LBP texture feature extraction and data broadening operations are added to the lung sound data cleaning to provide effective improvement of the accuracy of the model.(2)A single-channel convolutional long and short-term memory neural network(C NNLSTM)method is proposed to obtain lung sound data features serially.(3)A dual channel convolutional long and short term memory neural network(DCCLNN)method is proposed to acquire lung sound data features in parallel.In this paper,experiments are conducted on the R.A.L.E.? Lung Sounds dataset,and the current best Dalal＿CNN accuracy is 89.56%;while the accuracy of the CNNLSTM proposed in this study is 96.43% and the recognition accuracy of DCCLNN is 97.40%.From the experimental results,the accuracy of the CNNLSTM and DCCLNN methods proposed in this paper is improved by 6.87% and 7.84%,respectively,compared with that of Dalal＿CNN.