Heart sound is an acoustic wave generated by the mechanical movement of the heart and blood flow, and is a complicated, non-stationary signal composed of many signal sources. It can be divided into normal heart sounds and heart murmurs. Murmurs are abnormal signals that appear over wider ranges of frequency than normal heart sounds. They are generated at random spots in the whole period of heart sounds. The recognition of heart sounds is to differentiate heart murmurs through patterns that appear according to the generation time of murmurs. In this paper, a group of heart sounds was classified into normal heart sounds, pre-systolic murmurs, early systolic murmurs, late systolic murmurs, early diastolic murmurs, and continuous murmurs. The suggested algorithm was standardized by re-sampling and then added as an input to the neural network through wavelet transform. The neural network used Error Back - Propagation algorithm, which is a representative learning method, and controlled the number of hidden layers and the learning rate for optimal construction of networks. As a result of recognition, the suggested algorithm obtained a higher recognition rate than that of existing research methods. The best result was obtained with the average of 88% of the recognition rate when it consisted of 15 hidden layers. The suggested algorithm was considered effective for the recognition of automatic diagnosis of heart sound recognition programs.

Recently many researches concerning heart sound analysis are being processed with development of digital signal processing and electronic components. But there are few researches about recognition of heart sound, especially full cardiac cycled heart sound. In this paper, three new recognition methods about full cardiac cycled heart sound were proposed. The first method recognizes the characteristics of heart sound by integrating important peaks and analyzing statistical variables in time domain. The second method builds a database by principal components analysis on training heart sound set in time domain. This database is used to recognize new input of heart sound. The third method builds the same sort of the database not in time domain but in time-frequency domain. We classify the heart sounds into seven classes such as normal (NO) class, pre-systolic murmur (PS) class, early systolic murmur (ES) class, late systolic murmur (LS) class, early diastolic murmur (ED) class, late diastolic murmur (LD) class and continuous murmur (CM) class. As a result, we could verify that the third method is better efficient to recognize the characteristics of heart sound than the others and also than any precedent research. The recognition rates of the third method are 100% for NO, 80% for PS and ES, 67% for LS, 93 for ED, 80% for LD and 30% for CM.