Detection And Classification Of Cough Events Based On Audio Analysis Technology

Pneumonia and asthma are common symptoms of most respiratory diseases.It is difficult to differentiate these diseases without appropriate diagnostic tools and methods.Cough sounds may carry important disease-related information,and It is necessary to detect and differentiate disease carried by cough sounds to estimate the severity of a suspected patient’s condition.Therefore,in view of the large workload caused by traditional disease detection and other problems,the method in this paper focuses on cough analysis technology as a potential low-cost diagnostic tool,and proposes a general cough sound analysis framework,which provides important information for disease screening.theoretical value.The specific research contents are as follows:(1)Pathological Cough Detection Based on Mel Spectrogram and Improved Res Net-50 Residual Attention NetworkAn improved Res Net-50 residual attention network model is proposed for the detection of pathological coughs to achieve automatic classification of normal and pathological cough sounds.First,the obtained samples are preprocessed,and the Mel spectrogram is used as the input feature;secondly,the Res Net-50 model is improved by adding spatial and channel dual attention modules.The improved model can utilize the self-attention ability between different scales,so that the features focus on the most discriminative part of the cough feature,so as to extract deeper features.In addition,wavelet de-noising technology is used for the collected raw cough data to extract as much pure cough sounds as possible from the noisy signal.Experiments show that the improved algorithm can more accurately discriminate pathological cough,and the detection accuracy reaches 91.5%.(2)Automatic classification of asthma and pneumonia cough sounds esearch on Classification Algorithms of Cough Sounds.A classification algorithm for pneumonia and asthma cough diseases is proposed to automatically discriminate the pathological cough sounds of the two diseases.Based on the difference in time-frequency characteristics of the two types of cough signals,a feature extraction method using a combination of Nonnegative Matrix Factorization(NMF)and Short-Term Energy(STE)is proposed,reflecting from different perspectives.The essential characteristics of cough events are obtained,and dimensionality reduction processing is realized on this basis.The SVM model shows excellent sequence modeling ability for small sample sets.The nonlinear SVM with radial basis(RBF)as the kernel function is selected,and the key parameters of the model are optimized by particle swarm algorithm(PSO).For the asthma-cough recognition experiment,the accuracy,specificity and sensitivity were used as the main evaluation criteria,and the optimal threshold was obtained according to the task requirements.Through the calculation of Youden coefficient,the final threshold=0.7was obtained,and the recognition accuracy of asthma and cough was 92.8%.The experiments focus on the automatic classification problem of pneumonia and asthmatic cough,and the overall classification accuracy reaches 92.3%.