| 张涛,高新意,唐伟,丁碧云.基于神经网络的玻璃缺陷声学检测方法[J].声学技术,2018,37(5):488~495 |
| 基于神经网络的玻璃缺陷声学检测方法 |
| Acoustic detection method of glass defects based on neural network |
| 投稿时间：2017-08-22 修订日期：2017-10-18 |
| DOI：10.16300/j.cnki.1000-3630.2018.05.014 |
| 中文关键词: 缺陷检测 后向传播神经网络 特征提取 声学特征量 特征选择 |
| 英文关键词: defect detection Back Propagation(BP) neural network feature extraction acoustic feature feature selection |
| 基金项目:国家自然科学基金（61179045）、华为高校合作基金（2016120024000202）资助项目 |
|
| 摘要点击次数: 1017 |
| 全文下载次数: 661 |
| 中文摘要: |
| 描述了一种通过声学信号检测玻璃制品缺陷的方法。在实现步骤上，首先采集了不同缺陷类型的玻璃瓶敲击声，然后经过频谱变换及小波包变换，将敲击信号映射至不同的变换域中，并在每个变换域中提取信号的特征，从而将样本的缺陷信息对应为统计特征和物理特征，并采用基于互信息量的特征选择算法对特征空间进行降维；降维后的特征子集作为后向传播神经网络的输入参数，再由该神经网络实现对玻璃缺陷的自动化检测。结果表明，在已有实验样本数据下，该缺陷检测算法能准确高效地检测出存在缺陷的样本，识别结果的F-值稳定在95%左右。 |
| 英文摘要: |
| This paper describes the acoustic method to detect the defects of glass products. Firstly, the percussion signals of glass samples with different types of defects are collected. After the Fourier transform and the wavelet package transform, the signals are represented in different transform domains. The defects are expressed as statistical and physical features extracted from these domains. Secondly, a feature selection algorithm based on mutual information is used to reduce the number of dimensions of the feature space. The reduced feature subset is used as the input of BP neural network to realize the automatic detection of glass defects. The experimental results show that the proposed defect detection algorithm can identify defects well, and the F-measure is stable at about 95% in existing experimental samples. |
|
查看全文
查看/发表评论 下载PDF阅读器 |
| 关闭 |
