| Cough is a common symptom of respiratory diseases.Detection of cough sound can provide valuable reference for clinical diagnosis of respiratory diseases.At present,Deep Learning has become the mainstream method of speech recognition,and mixed structure Deep Learning network has achieved high recognition accuracy when applied to cough recognition in some special occasions.However,the application scenario in this thesis is 24-hour continuous cough monitoring to record the time,frequency and intensity of suspected cough.It requires the acquisition device to be portable and have a strong battery life,so it is not suitable to adopt a computation-intensive algorithm,nor can it frequently communicate with the cloud platform to obtain real-time recognition results.Therefore,this thesis chooses the traditional machine learning model to design the cough recognition algorithm,and tries to reduce the feature dimension,so as to reduce the calculation of the algorithm as much as possible on the premise of ensuring high recognition accuracy.The key of traditional machine learning is to select appropriate features.According to the generation mechanism of cough sound,cough sound is regarded as the result of modulation of vocal tract to excitation signal,and the features of vocal tract model and excitation signal energy are extracted to recognize cough.The main work is as follows:(1)Mel Frequency Cepstral Coefficients(MFCC)of the whole cough signal are often calculated in frames as the main feature of cough recognition.On this basis,MFCC of cough sound’s outbreak stage is selected as the feature and its rationality is analyzed.Meanwhile,the dimension of MFCC is determined.(2)The Linear Predictive Coding is used to build a vocal tract model,and then a filter is constructed to separate the vocal tract and the excitation signal,so as to suppress the influence of the modulation of the vocal tract and calculate the intensity of the excitation signal more accurately;Short-term energy is used to quantify the intensity of excitation signals,and Principal Component Analysis(PCA)is used to construct feature transformation matrix to extract the energy feature of cough sound excitation signals.The rationality of energy feature is analyzed with the characteristics of cough air flow.(3)Linear Support Vector Machine is used to compare the recognition effects of different feature parameters to further verify the validity of the features;The recognition model of cough sound is established based on Support Vector Machine with Gaussian kernel function,and the key parameters of Gaussian kernel function are determined.Then develop the cough sound automatic continuous detection software.The test results show that the sensitivity and specificity of the cough sound recognition model reach 93.64% and 97.50%,which lays a good foundation for the long-term cough monitoring and initial screening of suspected cough.The Developed Android application also basically meets the functional requirements and achieve the expected effect. |
