北京邮电大学学报

  • EI核心期刊

北京邮电大学学报 ›› 2020, Vol. 43 ›› Issue (5): 112-117.doi: 10.13190/j.jbupt.2020-075

• 研究报告 • 上一篇    下一篇

一种基于EEMD的异常声音识别方法

韦娟1, 顾兴权1, 宁方立2   

  1. 1. 西安电子科技大学 通信工程学院, 西安 710071;
    2. 西北工业大学 机电学院, 西安 710072
  • 收稿日期:2020-06-25 发布日期:2021-03-11
  • 作者简介:韦娟(1973-),女,副教授,E-mail:weijuan@xidian.edu.cn.
  • 基金资助:
    国家自然科学基金项目(51675425);陕西省重点研发计划项目(2018GY-181,2020ZDLGY06-09)

An Abnormal Sound Recognition Method Based on EEMD

WEI Juan1, GU Xing-quan1, NING Fang-li2   

  1. 1. School of Communication Engineering, Xidian University, Xi'an 710071, China;
    2. School of Mechanical Engineering, Northwestern Polytechnical University, Xi'an 710072, China
  • Received:2020-06-25 Published:2021-03-11

摘要: 为了优化组合特征在异常声音识别中的效率,提出一种用集合经验模态分解(EEMD)对异常声音帧信号进行有效性检测和提取多层特征的算法.首先对异常声音帧信号进行集合经验模态分解,得到固有模态函数;然后根据给定的固有模态函数层数阈值,对该帧信号进行有效性检测;再对有效帧信号的每一层固有模态函数提取梅尔频率倒谱系数、翻转梅尔频率倒谱系数、线性预测倒谱系数、短时能量和能量比,并将它们归一化后拼接成多层特征.根据提取的特征,用深度卷积神经网络实现异常声音识别分类.仿真结果表明,提出的新方法在4类异常声音识别中的识别率可以达到98.65%.

关键词: 异常声音识别, 集合经验模态分解, 多层特征, 深度卷积神经网络

Abstract: In order to optimize the efficiency of combined features in abnormal sound recognition,an algorithm for detecting the effectiveness of abnormal sound frame signals and extracting multi-layer features using ensemble empirical mode decomposition (EEMD) is proposed. Firstly,an ensemble empirical mode decomposition is performed on the abnormal sound frame signal to obtain the intrinsic model function,and then the validity of the frame signal is tested according to the given layer threshold of the intrinsic modal function. Finally,the Mel frequency cepstral coefficients,the inverted Mel frequency cepstral coefficients,the linear prediction cepstral coefficients,the short-time energy and energy ratio are extracted for each layer of the intrinsic modal function of the effective frame signal,and then all of them are normalized and spliced into multi-layer feature. According to the extracted features,the deep convolutional neural network is used to realize the classification and recognition of abnormal sound. Simulations show that the proposed new method can achieve a recognition rate of 98.65% in four types of abnormal sound recognition.

Key words: abnormal sound recognition, ensemble empirical mode decomposition, multi-layer feature, deep convolutional neural network

中图分类号: