Research On Classification Methods Of Pulmonary Adventitious Respiratory Sound Based On Parallel Encoders

The detection of pulmonary adventitious respiratory sound is essential in the early diagnosis of respiratory diseases,which can assist doctors to evaluate the pulmonary pathological tisuuse state of patients.The existing research methods only analyze from the perspective of single feature or statistical features,and do not mine the internal structure information of unlabeled data,so they are not well suitable for the detection of abnormal respiratory sound.To solve these problems,this paper proposes classification methods of pulmonary adventitious respiratory sound based on the parallel encoders.The main research works are as follows:Firstly,aming at the problems of insufficient single feature representation ability and interference of representation information irrelevant to the detection task,a classification method of pulmonary adventitious respiratory sound based on parallel encoders with attention is proposed.To mine the complementary information of two kinds of spectral features of respiratory sound,Mel-spectrogram and Mel-frequency cepstral coefficients,a group of parallel encoders are designed to encode the resulting features and fuse them according to the way of channel concatenation at the same time,so as to improve the classification performance of respiratory sound.Considering the discrepancy between features,not all the features extracted from the middle layer of the model are conducive to respiratory sound detection.Therefore,this paper constructs residual attention and channelspatial attention mechanism to mine the weights bettwen different features,so as to strengthen the information related to respiratory sound detection and further improve the performance.Experimental results on ICBHI 2017 dataset show that this method can effectively identify adventitious respiratory sound in three classification tasks.This method achieves the accuracy of 80.0% and 92.4% repectively in two types of two classification tasks,and the evaluation score of 56.76% in four classification task.Secondly,for the realistic challenges of high heterogeneity of pulmonary respiratory sound and relatively small amount of data,most methods only study from the aspects of spectrum analysis and model design,do not mine the internal potential structure information of unlabeled data,and can not accurately detect adventitious respiratory sound.Therefore,a self-supervised contrastive classification method of pulmonary adventitious respiratory sound based momentum encoder is proposed in this paper.The method consisits of two stages: self-supervised constrastive pre-training and fully supervised fine-tuning classification.In the self-supervised contrastive pre-training stage,the internal information structure of unlabeled data samples is mined by parallel encoders and the representation consistency contrastive loss is calculated to train the model to achieve the purpose of self-supervised learning.After that,the encoder obtained from the self-supervised contrastive pre-training stage is transferred to the fully supervised classification stage as the classifier.After fine-tuning through the labeled data,the pulmonary respiratory sound is classified.Experimental results on ICBHI 2017 dataset show that this proposed method can effectively improve the classification performance.This method achieves the accuracy of 82.6% and 92.8% respectively in two types of two classification tasks,and the evaluation score of 59.61% in four classification task.Meanwhile,experimental results also show that self-supervised pre-training can effectively improve the performance of the existing classifiers.