Study On Lung Sound Recognition Algorithm For High Altitude Pulmonary Edema

High altitude pulmonary edema(HAPE),as one of the common mountain diseases in the plateau area,has an acute onset and rapid development,which poses a serious threat to people’s life and health.However,if it can be found in time and treated in time,it can greatly reduce the harm to the human body.Lung sound auscultation is one of the common methods for the diagnosis of high altitude pulmonary edema.Through the analysis of lung sounds in different parts of the human body,it can quickly determine whether the patient is suffering from high altitude pulmonary edema and the stage of the disease,so as to timely prescribe drugs and protect people’s life and health.However,the diagnosis of lung sounds requires professional medical personnel.If the intelligent recognition of lung sounds and rapid and accurate diagnosis of lung sounds can be realized,it can not only assist doctors in the analysis of the disease,but also monitor the lung sounds,timely warn people entering the plateau area,and reduce the harm of high altitude pulmonary edema to the human body.The lung sounds in this paper are from the ICBHI 2017 database.The lung sound in the database is composed of the lung sound cycles of multiple segments and multiple labels,and the lung sound data with a single label is obtained after cutting the lung sound according to the label.The obtained lung sound data were removed by FIR filter and wavelet threshold noise reduction to remove the interference of heart sound,and relatively pure lung sound was obtained.The spectral features of lung sounds were obtained by Mel spectrogram(Mel),short time Fourier transform(STFT),wavelet transform(WT)and constant Q transform(CQT),which made the lung sounds visualized.LBP is used to enhance the texture features of the extracted feature spectrograms,and MMixup is used to expand the data to generate virtual samples and improve the situation of data imbalance.To prepare for the subsequent use of convolutional neural network for intelligent classification of lung sounds.The single-input CNN model and the dual-input Dual CNN model with fusion features as input are designed and built.The original random data set and the augmented data set are divided by 85% training set and 15% test set.The training set is to train two CNN models.After learning the characteristics of lung sounds,the test set is used to simulate new samples to test the lung sound classification effect of the model.The test results were evaluated in five aspects: confusion matrix,accuracy,sensitivity,specificity and ICBHI score.Finally,WT and Mel fusion features were used to achieve the best effect in the dual-input Dual CNN model,with an accuracy of 93.59%,sensitivity,specificity and ICBHI scores of 93.96%,92.3% and93.13%,respectively.Each score is above 92%,and the model has a good lung sound classification effect.