基于EEMD的异常声音多类识别算法

Abnormal sound multiclass recognition algorithm based on EEMD

为了提高异常声音信号的识别率,提出一种将总体平均经验模态分解和梅尔频率倒谱系数、短时能量及能量比相结合的特征提取改进算法,并对决策导向无环图支持向量机多类识别算法进行改进.首先对声音信号进行分帧,然后对每帧信号进行总体平均经验模态分解得到固有模态函数,最后对每层固有模态函数提取梅尔频率倒谱系数、短时能量和能量比特征.根据提取的特征,采用改进的决策导向无环图支持向量机算法对五种异常声音信号进行识别.仿真结果表明：改进的特征提取算法和决策导向无环图支持向量机多类识别算法相比改进前识的别率分别提高了2%和2.5%.

In order to improve the recognition rate of abnormal sound signals,the improved feature extraction algorithm was proposed,combining ensemble empirical mode decomposition（EEMD） with Mel-frequency cepstral coefficient,short-time energy and energy ratio.Meanwhile,the decision directed acyclic graph support vector machine algorithm was improved.First,frame processing was made for the sound signals,and then intrinsic mode functions were obtained for each frame signal by using EEMD.Finally,the features of Mel-frequency cepstral coefficient,short-time energy and energy ratio were extracted from each layer of intrinsic mode functions.According to the extracted features,five kinds of abnormal sound signals were recognized by using improved decision directed acyclic graph support vector machine algorithm.Results show that the recognition rates of abnormal sound signals using improved feature extraction algorithm and decision directed acyclic graph support vector machine multiclass recognition algorithm are increased by 2% and 2.5% respectively than that using former algorithm without improvement.