基于MLFCC特征和P-CRNN的心音分类研究

Research on Heart Sound Classification Based on Mixed Features of MFCC and LFCC and Parallel Convolutional Recurrent Neural Network

摘要: 心音分类是医学领域中一个重要的诊断任务。但由于数据量不足、宽频带信号表征能力欠缺、心音序列上下文学习存在挑战等问题，该领域依旧存在许多的不足之处。针对以上问题，提出了一种心音分类的新方法：对每段心音信号进行重叠切片，截取成2s的信号作为样本；随后采用改进的梅尔频率倒谱系数(MFCC)与线性频率倒谱系数(LFCC)分别提取心音信号相应频率系数，并分别计算其一阶差分作为融合特征。分类网络使用提出的并行卷积递归神经网络(P-CRNN)方法进行训练。实验表明，相比于其他传统识别方法，所提方法对心音信号分类有明显提高。

Abstract: Heart sound classification is a crucial diagnostic task in the medical field. However, due to the insufficient amount of data, the lack of wideband signal representation ability, and the challenges of contextual learning of heart sound sequences, there are still many shortcomings in this field. To address these issues, a novel method for heart sound classification was proposed. In this method, each signal was segmented into 2-second overlapping slices, and the corresponding frequency coefficients were extracted using improved Mel Frequency Cepstral Coefficients (MFCC) and Linear Frequency Cepstral Coefficients (LFCC). The first-order difference was calculated as fusion features, and the proposed Parallel Convolutional Recurrent Neural Network (P-CRNN) method was trained for classification in comparison to traditional recognition methods. The experiment shows that compared to other traditional recognition methods, the proposed method has a significant improvement in heart sound signal classification.