Research On Lung Sound Classification Algorithm Based On Parallel Pooled CNN

Pathological lung sound is an important index for diagnosing lung health.Fully analyzing the characteristics of pathological lung sound and classifying and identifying the characteristics is an effective means to detect abnormal lung conditions and prevent further deterioration of the disease.Pathological lung sound belongs to the aperiodic signal with low frequency and low amplitude.Due to its complex generation mechanism and various acquisition positions and methods,the components of the signal are complex and it is difficult to identify and classify.Therefore,professional medical staff are needed to accurately judge the type of pathological pulmonary sound in the signal.In order to provide an effective reference for medical professionals in diagnosing lung health conditions and reduce the workload of staff,more and more studies have begun to use machine learning to intelligently recognize and classify lung sound.This paper analyzes the lung sound samples in the open source lung sound database（ICBHI 2017）provided by the International Conference on biomedical and health informatics,and designs FIR filtering and wavelet denoising for environmental and heart sound interference.Mixup and LBP were used to enhance the lung sound samples.In this paper,a lung sound recognition algorithm based on convolution neural network is established.For the classification and recognition of two typical pathological lung sounds,blister sound and wheezing sound,four spectral feature extraction methods based on STFT transform,wavelet transform,Mel spectrum and constant Q transform are proposed.The characteristics of sonogram were analyzed and described,which provided image basis for the subsequent identification and classification of pathological pulmonary sound.In the construction of convolutional neural network,in order to achieve high classification accuracy,this paper constructs a bi CNN-SVM hybrid network model that can learn two different features at the same time.The parallel pool structure CNN is used for depth feature extraction,and the rsm-svm integrated learning classifier outputs the classification results.ICBHI 2017 was used to train and test the model,and the effectiveness of bi CNN-SVM model was evaluated from the aspects of accuracy,confusion matrix,F1 score and official score.When the ICBHI training set and test set are divided by using the ratio of 85:15,the accuracy of parallel pooled CNN and biCNN-SVM hybrid models are 63.3%and 69.8% respectively.