正则化深度神经网络的防腐层缺陷超声信号识别
Ultrasonic signal recognition of anticorrosive coating defects based on regularized deep neural network
沈显庆 1王童正 1张腾鹤 1徐梦泽1
作者信息
- 1. 黑龙江科技大学 电气与控制工程学院,哈尔滨 150022
- 折叠
摘要
传统智能算法识别防腐层缺陷超声信号依赖于特定技术来提取特征,过程复杂不适用于工业应用,在识别混合频率数据时精度不佳.通过建立铝板防腐层缺陷混合频率的超声信号数据集,训练基于Drop out正则化的深度神经网络DNN,识别脱粘、弱粘接、斑点型脱粘和防腐层结构缺陷.结果表明,在混合频率数据集上给定的DNN结构实现了 92.76%的最高精度,比常用的单隐层神经网络提高了13.87%,在不从超声信号中人工提取任何特征的情况下,高精度地识别了缺陷.
Abstract
This paper intends to address the problem that traditional intelligent algorithms for identif-ying ultrasonic signals of anti-corrosion coating defects rely on specific techniques to extract features with complexed process not for industrial applications and poor accuracy in identifying mixed frequency data.The study involves establishing an ultrasonic signal dataset for the mixed frequencies of defects in the an-ti-corrosion layer of aluminum plates;and raining a deep neural network DNN based on Drop out regulari-zation with the defects types of partial debonding,weak bonding,spot type debonding,and structural de-fects in the anti-corrosion layer.The results show that the given DNN structure achieves the highest accu-racy of 92.76%in mixed frequency datasets,which is 13.87%higher than the common single hidden layer neural networks.It can accurately identify defects without manually extracting any features from ul-trasound signals.
关键词
防腐层缺陷识别/Drop/out正则化/深度神经网络Key words
anti-corrosion coatings flaws classification/Drop out regularization/deep neural network引用本文复制引用
出版年
2024
国家科技期刊平台
NSTL
万方数据