Page 61 - 《应用声学》2023年第1期

第 42 卷 第 1 期 Vol. 42, No. 1
2023 年 1 月 Journal of Applied Acoustics January, 2023

⋄ 研究报告 ⋄

基于时频分帧能量熵的陶瓷制品敲击声波

信号特征识别 ∗

刘利平 1,2 蒋柳成 1† 乔乐乐 1 孙 建 1 高世妍 1

(1 华北理工大学人工智能学院 唐山 063210)
(2 华北理工大学矿业工程学院 唐山 063210)
摘要：针对现有陶瓷制品敲击声波信号特征提取方法中提取的特征代表性降低的问题，该文提出结合最大
重叠离散小波包变换 (MODWPT) 和时频分帧能量熵的特征提取方法。首先采用 MODWPT 将信号分解为
4 层，再对每个节点的子信号分帧后计算各个节点的时频分帧能量熵，然后根据能量分布特征选择了前 6 个
节点的时频分帧能量熵特征，最后构建随机森林分类器完成识别。将该方法和 MODWPT 时频分段能量熵、
MODWPT 归一化能量特征两种方法进行比较。实验结果表明，相比 MODWPT 时频分段能量熵、MODWPT
归一化能量两种特征提取方法，MODWPT 时频分帧能量熵能提升特征的代表性，具有更优的陶瓷制品敲击
声波信号特征识别性能，其识别的 F 1 值达到了 98.46%，相比上述两种方法分别提升 F 1 值 3.22%、1.86%。
关键词：敲击法；最大重叠离散小波包变换；时频分帧能量熵；模式识别
中图法分类号: TB52 文献标识码: A 文章编号: 1000-310X(2023)01-0057-10
DOI: 10.11684/j.issn.1000-310X.2023.01.008

Coin-tap sound signal characteristics recognition of ceramic products based on

time-frequency framing energy entropy

LIU Liping 1,2 JIANG Liucheng 1 QIAO Lele 1 SUN Jian 1 GAO Shiyan 1

(1 College of Artificial Intelligence, North China University of Science and Technology, Tangshan 063210, China)
(2 College of Mining Engineering, North China University of Science and Technology, Tangshan 063210, China)

Abstract: In view of the problems of decreasing of extracted feature representativeness in feature extraction
method of ceramic products coin-tap sound signal characteristics recognition, the feature extraction method
combining maximum overlap discrete wavelet transformation (MODWPT) with time-frequency framing energy
entropy is proposed. Firstly, the MODWPT method is applied to decompose the coin-tap sound signal to
level 4. Then the time-frequency framing energy entropy of each node is calculated after sub-signal framing
of each node. And then the time-frequency framing energy entropy features of the first 6 nodes are selected
according to the energy distribution feature. Finally, the random forest (RF) classifier is constructed to
complete the identification. This method is compared with two methods of MODWPT time-frequency segment
energy entropy and MODWPT normalized energy features. The experimental results show that compared
with MODWPT time-frequency segment energy entropy and MODWPT normalized energy, MODWPT
time-frequency framing energy entropy can improve the representativeness of extracted feature, and has better

2021-11-26 收稿; 2022-02-16 定稿
河北省省级科技计划项目 (20327218D), 华北理工大学研究生创新项目 (2019B28)
∗
作者简介: 刘利平 (1977– ), 女, 河北唐山人, 博士研究生, 研究方向: 矿物材料、模式识别及智能控制。
† 通信作者 E-mail: 3448599779@qq.com