基于欠定盲源分离理论与深度学习的声音样本集获取与分类方法

Acquisition and Classification of Sound Samples Based on Underdetermined Blind Source Separation Theory and Depth Learning

针对现有盲源分离算法在单一传感器条件下分类精度不足的问题,首先提出了一种K-DPC聚类新型欠定盲源分离算法,得到了混合噪声信号中的各声源信号分量,构建了独立声源信号库;借助k-UBSS算法中的混合估计矩阵,提出了一种利用独立声源信号库获得大量混合时频谱图的方法,最后分别通过VGGNet-16和ResNet-50模型进行训练,实现了混合声音信号的分类。对比测试结果表明:k-UBSS算法的偏角差均小于2,归一化均方误差达-38.372,较现有K-means和DPC算法的精度有很高提升,VGGNet-16和ResNet-50分类准确率分别可达到93.75%和99.2%。仿真结果验证了K-UBSS算法的准确性,以及混合时频谱图获取方法的可行性,实现了单一传感器下源声音信号的快速分类。采用该算法能有效解决深度学习时样本不足的问题,并对噪声采集和治理具有重要意义。

The existing blind source separation algorithms with single sensor is insufficient in calculating classification accuracy.Considering that,this paper proposed a new K-DPC clustering underdetermined blind source separation algorithm to obtain each sound source signal component in mixed noise signal and an independent sound source signal library.Based on the hybrid estimation matrix in the k-UBSS algorithm,we proposed to obtain mass mixed time spectrum charts using independent sound source signal libraries.Finally,we performed training by VGGNet-16 and ResNet-50 models and achieved the classification of mixed sound signals.The comparison test results show that the deflection angle differences of the k-UBSS algorithms are all less than 2,and that the normalized mean square error is-38.372,which is much higher than that of the existing K-means and DPC algorithms.VGGNet-16 And ResNet-50 classification accuracy reach 93.75%and 99.2%respectively.The simulation results verify the accuracy of the K-UBSS algorithm and the feasibility of the hybrid time-frequency spectrum acquisition method,which realizes the rapid classification of the source sound signal under a single sensor.The proposed algorithm provides sufficient samples for deep learning and is of great significance for noise collection and governance.