Research On Time-Frequency Energy Spectrogram Analysis Based On Wigner-Ville Distribution And Classification Algorithm Of Heart Sound Signal With Murmurs

Heart sound is one of the most important human physiological signals.Auscultation is not only simple, non-invasive but also can detect abnormal earlier.Because of the complexity and non-stationary of heart sound signal, using moderndigital signal processing methods for analysis and reorganization has become theindispensable means to learn the status of cardiovascular. The thesis studied andanalyzed the heart sound and murmurs based on the aim of clinical diagnosis needs,including the formation mechanism of the heart murmurs, the chosen of signal featurevectors, the classification of normal heart sound and four types of pathological heartmurmurs (aortic stenosis, aortic regurgitation, mitral regurgitation and pulmonarystenosis).The work of the thesis includes the following parts:1) The characteristics analysis of normal heart sound and heart murmurs.The formation mechanism and time-domain waveform characteristics are studied aswell as the AR model power spectrum estimation, providing basic information onfrequency domain and reference for the choosing of feature vector.2) Multi-components separation algorithm of heart murmurs. Singularspectrum method based on principal component analysis was applied for separating thenormal heart sound components and murmur components by the processes of singularvalue decomposition and reconstruction. The result showed that the multi-componentseparation methods can effectively inhibit the cross-term interference in Wignerdistribution.3) The analysis methods of time-frequency energy spectrogram for normal heartsound and murmurs. Several signal time-frequency analysis methods were compared:short-time Fourier transform, wavelet transform and the Wigner distribution. Theresolution of STFT spectrogram was greatly influenced by the width of the windowfunction. The wavelet transform can get the scalogram but the selection of base waveletand scale parameter is not easy. As a result, Wigner distribution was chosen fortime-frequency analysis of heart sounds and murmurs. The two dimensionaltime-frequency energy spectrogram with high resolution can well reflect thecharacteristics of the signal in time-frequency domain and energy domain.4) The feature extraction of heart sounds and murmurs. Normal heart sounds and four types of heart murmurs from3M Littmann Stethoscopes database were processedby the proposed methods. Then the features from time domain, frequency domain andenergy such as the murmurs’ duration, peak frequency and energy fraction wereextracted as the feature vector, which is necessary for classification analysis.5) The classification method of hear sound and murmurs. The support vectormachine was chosen as the classifier because of the limited samples of pathologicalheart murmurs. The selection of kernel function and the module for multi-classes ofsamples were studied. The grid optimization method using the classification accuracy asthe criterion was selected to determine the optimal values of kernel function and therelaxation variable. Then the classification model of support vector machine adapted forheart sounds and murmurs was established based on the optimal values. Normal heartsound and four types of pathological heart murmurs(aortic stenosis, aortic regurgitation,mitral regurgitation and pulmonary stenosis），each type selected30cases forlearning set and10cases for testing set. The results showed that the classificationaccuracy is higher than90%on average which proved the validity of the proposedalgorithm.