基于改进的MFCC与CNN心音信号识别方法的研究

Recognition and classification of heart sound signals based on LMFP and CNN

摘要:

心音分类在心血管疾病的早期检测中起着至关重要的作用,特别是对小型初级卫生保健诊所、缺少专业人员陪护的家庭等检测。为提高心音信号数据类别间的可辨别性,进一步提高分类精度,提出了一种基于多预处理法(LMFP)和卷积神经网络(CNN)模型的分类方法。首先,原始数据的采集频率为44100Hz,所处理数据量比较大,需要对数据下采样处理,以减少不必要的数据量。第二,分别采用带通滤波器、SG滤波器与MFCC预处理,提取心音数据特征,并将一维数据转换为二维数据或者图谱,并计算数据PCA变换矩阵。第三,将预处理后的二维数据对应的PCA变换矩阵相乘,这是LMFP的主要部分,可减少不必要的维数,使数据更具代表性。最后,将处理后的数据,输入到本文的模型CNN中。为了验证LMFP+CNN算法的有效性和可靠性,利用PASCAL挑战数据部分数据集进行了实验。通过与其他方法、卷积神经网络不同层数的比较,证明了该方法的优越性。实验结果表明,本文提出的方法可有效达到97.21%的准确率。

Abstract:

Heart sound classification plays a crucial role in the early detection of cardiovascular disease, especially in small primary health care clinics and in homes without a professional presence. In order to improve the discriminability between categories of heart sound signal data and further improve the classification accuracy, a classification method based on multi-preprocessing method (LMFP) and convolutional neural network (CNN) model was proposed. First of all, the acquisition frequency of the original data is 44100Hz, and the amount of data processed is relatively large, so it is necessary to downsample the data to reduce the unnecessary amount of data.? Second, bandpass filter, SG filter and MFCC were used for preprocessing to extract the features of heart sound data, and the one-dimensional data was converted into two-dimensional data or spectrogram, and the data PCA transformation matrix was calculated. Third, multiply the PCA transformation matrix corresponding to the preprocessed two-dimensional data, which is the main part of LMFP. It can reduce unnecessary dimensions and make the data more representative. Finally, the processed data is input into the model CNN of this paper. In order to verify the validity and reliability of LMFP+CNN algorithm, experiments were carried out with partial data set of PASCAL challenge data. Compared with other methods and convolutional neural networks with different layers, the superiority of this method is proved. The experimental results show that the method proposed in this paper can effectively achieve 97.21% accuracy.