改进的ESMD用于公共场所异常声音特征提取

由于公共场所异常声音的特殊性及背景噪声的复杂性,极点对称模态分解(ESMD)用于异常声音分解时,存在一些理论和技术上的缺陷。经分析认为公共场所异常声音为非线性、非平稳信号,背景噪声服从T分布。为此,提出改进的ESMD用于公共场所异常声音分解,得到有利于识别的特征。所提出方法的特点是将T分布噪声序列添加到具有背景噪声的异常声音信号中,以减小背景噪声对特征提取的影响;将模态分量的排列熵作为判定异常声音与背景噪声的准则,自适应筛选有效的模态分量;用对称中点插值法替代极值中点奇偶插值法,以缓解ESMD插值端点不明确带来的模态失真。在公共场所异常声音数据库上进行了相关实验。实验结果表明,所提出的方法与目前典型的时频信号处理方法相比,在提高公共场所异常声音分类识别率的同时,缩短异常声音的分解时间,是一种有效的公共场所异常声音特征提取方法。

Developed ESMD for the feature extraction of abnormal sound in public places

Because of the particularity of the abnormal sound signal and the complexity of background noise in public places, when the extreme point symmetric mode decomposition (ESMD) is used as the decomposition tool for abnormal sound signal, there exist some inherent theoretical and technical drawbacks. Through analysis, this paper concludes that the abnormal sound signal is a nonlinear, non stable signal and the background noise in public places obeys T distribution. Hence, a developed ESMD (D ESMD) method is proposed and used for the abnormal sound signal decomposition, and then the features of the abnormal sound signal in public places are extracted for identification. The features of the D ESMD method are adding T distribution noise into the abnormal sound signal with background noise to reduce the influence of background noise on features extraction, and the permutation entropy of the mode component is taken as the criterion for distinguishing abnormal sound signal and background noise, the effective mode components are selected adaptively. The symmetry extreme center interpolation method is used to replace the extreme center odd even interpolation method to overcome the mode mixing of ESMD caused by the ambiguity of mode endpoint. Related experiments were conducted on the abnormal sound database of public places. The results demonstrate that the proposed method can shorten the decomposition time for the abnormal sound signal, while improve the classification identification rate for the abnormal sound signal of public places compared with current typical time frequency signal processing methods, and is an effective extraction method for the abnormal sound features of public places.