基于时频重排的梅尔频率倒谱系数分类算法研究OA

Mel frequency cepstral coefficients(MFCC)can effectively characterize heart murmurs by mimicking human auditory char-acteristics,but they cannot achieve a good time-frequency resolution.To address the issue that MFCC struggles to accurately capture and depict the subtle features of transient events,leading to limited representation capability for such events,a Mel spectrogram optimization al-gorithm for heart sound classification based on Convolutional Neural Networks(CNN)after time-frequency rearrangement is proposed.First,heart sounds are preprocessed through framing and denoising;then,Short-Time Fourier Transform(STFT)is applied to the preprocessed heart sound signals to rearrange their time-frequency domain,extracting the rearranged Mel spectrogram.Finally,the convolutional neural network is applied to the Yaseen dataset,which includes normal and four types of abnormal heart sounds,achieving a classification accuracy of 99.0%.The study shows that the proposed method of optimizing Mel spectrograms after time-frequency rearrangement can significantly improve the CNN's accuracy in classifying heart sound signals,providing new insights for intelligent auscultation of heart sounds.