基于MFCC与GFCC混合特征的先心病心音分类研究

Research on Heart Sound Classification of CHD Based on MFCC and GFCC Mixed Features

为提高心音信号的分类准确率,提出一种基于梅尔频率倒谱系数与Gammatone频率倒谱系数的混合特征(MFCC与GFCC混合特征)的先心病心音信号分类算法。首先用db6小波双参数可调阈值函数对心音信号降噪,再用基于逻辑回归的隐半马尔可夫模型自动分段以提取单个心动周期;然后对信号加汉宁自卷积窗并提取心音的MFCC与GFCC混合特征,再用主成分分析法进行降维,以减少计算量;最后采用深度学习模型Inception v4进行分类识别,并与其它传统识别方法做了分类比较研究。用所提出的方法对1600例心音样本进行了分类测试,实验结果表明,上述方法对先心病心音的分类准确率比传统识别方法有明显提高,分类准确率达91.25%。

In order to improve the classification accuracy of PCG,a novel classification algorithm for CHD heart sound signal was put forward in this paper.It is based on the mixed features of Mel-frequency cepstral coefficients and Gammatone frequency cepstral coefficients(mixed features of MFCC and GFCC).First,the denoised heart sound was segmented to the cardiac cycle by using the db6 Wavelet two-parameter threshold function.Then,PCG was segmented into every single cardiac cycle accurately.The function of the Hanning self-convolution window was put to each segmentation of heart sound.After that,the mixed features of MFCC and GFCC were extracted.The principal component analysis was used to reduce the dimension of the extracted features.Finally,the deep learning model Inception v4 network was used to classify the normal and abnormal heart sound.1600 cases of heart sounds were used in this study.The results show that compared with other traditional recognition methods,the novel method has a significantly higher recognition rate for CHD heart sounds and the classification accuracy of 91.25%is achieved.