摘要
当前太赫兹(Terahertz,THz)厚度测量主要基于飞行时间理论,然而热障涂层微观结构不均匀,精确测厚需要引入频域信息在线提取折射率。为此,基于THz波传输特性,提出了关键时频信息融合方法,结合深度学习减小了反射峰变形对折射率测量精度影响,提高了微观结构不均匀条件下热障涂层测厚准确性。首先构建了考虑热障涂层表界面粗糙度THz信号解析模型,发现了测厚所需飞行时间与折射率可利用时域和频域数据求解。然后利用小波变换实现了THz信号时频融合,依据朗伯比尔定理揭示了完整时频图含有冗余信息物理机理。最后基于厚度特征获取了关键时频图,以此为输入,构建了卷积神经网络测厚方法。试验与仿真结果表明,所提出方法与反射峰提取、时域特征机器学习、时域信号、插值下采样时频图、完整时频图卷积神经网络测厚方法相比,平均绝对误差分别减小4.65μm、3.08μm、2.51μm、0.58μm、0.25μm,较输入完整时频图网络训练耗时降低1446 s。关键时频融合方法可提高热障涂层微观结构不均匀与反射峰变形条件下THz测厚精度。
The current Terahertz(THz)thickness measurement is mainly based on the theory of time-of-flight,but the accurate thickness measurement of thermal barrier coatings requires to introduce the THz frequency signal to extract the refractive index online.To this end,a fusion method of the key time-frequency information is presented by the transmission rules of THz waves,and the deep learning is integrated to decrease the thickness measurement errors,which reduces the influence of peak distortions for refractive index extractions and improves the accuracy of uneven thermal barrier coatings thickness measurements.To begin with,an analytical model of THz signals considering the roughness of thermal barrier coatings is constructed,and it is found that time-of-flight and refractive index can be solved by time-domain and frequency-domain data.Then,the wavelet transform is utilized to enable the time-frequency fusion of THz signals.In terms of Beer-Lambert’s law,the physical mechanism of the entire time-frequency image containing redundant information is revealed.Finally,according to the thickness feature,the key time-frequency image is acquired as the input of the convolutional neural network to establish a thickness measurement method.The simulated and experimental results demonstrate that compared with the other four thickness measurement approaches,including the peak locations,machine learning with time-domain signals-based features,convolutional neural networks with the temporal data,down-sampling and entire time-frequency images,the average absolute errors of the approach with time-frequency images are reduced by 4.65μm,3.08μm,2.51μm,0.58μm and 0.25μm,respectively.In contrast to the convolutional neural network utilizing the entire time-frequency image,the training time of the convolutional neural networks with the time-frequency image is decreased about 1446 s.The key time-frequency fusion method can be employed to improve the THz thickness measurement accuracy under the conditions of thermal barrier coatings with the uneven microstructure and peak distortions.
作者
孙凤山
范孟豹
曹丙花
刘林
SUN Fengshan;FAN Mengbao;CAO Binghua;LIU Lin(School of Mechatronic Engineering,China University of Mining and Technology,Xuzhou 221116;School of Information and Control Engineering,China University of Mining and Technology,Xuzhou 221116;Beijing Aerospace Institute for Metrology and Measurement Technology,Beijing 100076)
出处
《机械工程学报》
EI
CAS
CSCD
北大核心
2023年第14期10-22,共13页
Journal of Mechanical Engineering
基金
国家自然科学基金(62071471)
徐州市科技计划基础研究计划面上(KC21045)
江苏高校优势学科建设工程、中国矿业大学未来杰出人才助力计划(2022WLKXJ014)
江苏省研究生科研与实践创新计划(KYCX22_2510)资助项目。
关键词
太赫兹无损检测
热障涂层测厚
时频融合
解析模型
卷积神经网络
terahertz nondestructive testing
thermal barrier coating
time-frequency fusion
analytical model
convolutional neural network
