基于高阶累积量的参数化双谱分析的肺音特征提取

用高阶谱分析方法,对肺音信号进行了特征提取.以非高斯白噪声激励的AR(AutoRegressive)参数模型对肺音信号进行建模,导出了基于三阶累积量的三阶递推方法的非高斯参数化双谱的计算方法,用双谱的互相关估计模型阶次,并对肺音数据进行了参数化的双谱估计,给出了在双频域内从双谱及其切片谱提取肺音特征信息的方法,并利用该方法对正常、哮喘和细罗音3种肺音目标进行了神经网络的识别实验.结果表明:所提出的特征提取方法大大降低了输入特征的维数,具有较高的识别率

Lung sound feature extraction based on parametric bispectrum analysis of higher-order cumulants

The lung sound feature were extracted by using higher-order spectral technique. The paper presents the lung sound model using the AR parametric model excited by non-Gaussian white noise, and the nonGaussian parametric bispectrum algorithm was developed based on third-order recursion via third-order cumulants. The orders of the AR model were selected via the bispectral cross correlation and were applied to estimate the parametric bispectrum of the lung sound time series. The feature extraction method deriving from the bispectrum and the diagonal slices was also proposed in double frequency domain so that three kinds of lung sound signals (normal, asthma and fine) were investigated for pattern classification by artificial neural network(ANN). The results show that the new analysis method surprisingly decrease the number of ANN input vectors and highly improve the efficiency of recognizing lung sound.