利用稀疏自编码器网络对典型目标一维高分辨距离像(high resolution range profile,HRRP)进行了学习训练,基于各层权重系数矩阵定义了一种综合权重系数,通过综合权重系数和降维特征与散射中心特征的对比分析,发现稀疏自编码器深层特征...利用稀疏自编码器网络对典型目标一维高分辨距离像(high resolution range profile,HRRP)进行了学习训练,基于各层权重系数矩阵定义了一种综合权重系数,通过综合权重系数和降维特征与散射中心特征的对比分析,发现稀疏自编码器深层特征与散射中心特征之间具有一定的关联性,并对综合权重系数和深层降维特征的物理意义进行了解释。首先针对HRRP构建稀疏自编码器网络,经过深层学习后获取训练后的权重系数和降维后的特征,并与散射中心的位置特征和强度分布特征进行关联性分析。结果表明,综合权重系数矩阵为与散射中心密切相关的类字典系数矩阵,反映了距离域强散射中心位置随角度变化的可能的分子集;降维特征能够实现对强散射中心的学习和提取,反映了强散射中心位置和强度随角度的变化。最后分析了网络训练层数和降维维数对学习训练结果的影响,可指导后续网络参数的选择。文章首次针对雷达HRRP数据开展深度学习特征的可解释性研究,为后续深度学习在雷达数据处理中的广泛应用提供了有益的导引。展开更多
It is difficult to quantitatively detect defects by using the time domain or frequency domain features of Lamb wave signals due to their dispersion and multimodal characteristics.Therefore,it is important to discover ...It is difficult to quantitatively detect defects by using the time domain or frequency domain features of Lamb wave signals due to their dispersion and multimodal characteristics.Therefore,it is important to discover an intrinsical parameter of Lamb waves that could be used as a damage sensitive feature.In this paper,quantitative defect detection in aluminium plates is carried out by means of wavenumber analysis approach.The wavenumber of excited Lamb wave mode is a fixed value,given a frequency,a thickness and material properties of the target plate.When Lamb waves propagate to the structural discontinuity,new wavenumber components are created by abrupt wavefield change.The new wavenumber components can be identified in the frequency-wavenumber domain.To estimate spatially dependent wavenumber values,a short-space two-dimensional Fourier transform(FT)method is presented for processing wavefield data of Lamb waves.The results can be used to determine the location,size and depth of rectangular notch.The analysis techniques are demonstrated using simulation examples of an aluminium plate with a rectangular notch.Then,the wavenumber analysis method is applied to simulation data that are obtained through a range of notch depths and widths.The results are analyzed and rules of the technique with regards to estimating notch depth are determined.Based on simulation results,guidelines for using the technique are developed.Finally,experimental wavefield data are obtained in aluminium plates with rectangular notches by a full noncontact transceiving method,i.e.,laser-laser method.Band-pass filtering combined with continuous wavelet transform is used to extract a certain frequency component from the full laser-induced wavefield with wide band.Shortspace two-dimensional FT method is used for further processing full wavefield data at a certain frequency to estimate spatially dependent wavenumber values.The consistency of simulation and experimental results shows the effectiveness of proposed wavenumber method for quantitative rectangular notch detection.展开更多
文摘利用稀疏自编码器网络对典型目标一维高分辨距离像(high resolution range profile,HRRP)进行了学习训练,基于各层权重系数矩阵定义了一种综合权重系数,通过综合权重系数和降维特征与散射中心特征的对比分析,发现稀疏自编码器深层特征与散射中心特征之间具有一定的关联性,并对综合权重系数和深层降维特征的物理意义进行了解释。首先针对HRRP构建稀疏自编码器网络,经过深层学习后获取训练后的权重系数和降维后的特征,并与散射中心的位置特征和强度分布特征进行关联性分析。结果表明,综合权重系数矩阵为与散射中心密切相关的类字典系数矩阵,反映了距离域强散射中心位置随角度变化的可能的分子集;降维特征能够实现对强散射中心的学习和提取,反映了强散射中心位置和强度随角度的变化。最后分析了网络训练层数和降维维数对学习训练结果的影响,可指导后续网络参数的选择。文章首次针对雷达HRRP数据开展深度学习特征的可解释性研究,为后续深度学习在雷达数据处理中的广泛应用提供了有益的导引。
基金supported by the National Natural Science Foundation of China(Nos.51475012,11772014,and 11272021)
文摘It is difficult to quantitatively detect defects by using the time domain or frequency domain features of Lamb wave signals due to their dispersion and multimodal characteristics.Therefore,it is important to discover an intrinsical parameter of Lamb waves that could be used as a damage sensitive feature.In this paper,quantitative defect detection in aluminium plates is carried out by means of wavenumber analysis approach.The wavenumber of excited Lamb wave mode is a fixed value,given a frequency,a thickness and material properties of the target plate.When Lamb waves propagate to the structural discontinuity,new wavenumber components are created by abrupt wavefield change.The new wavenumber components can be identified in the frequency-wavenumber domain.To estimate spatially dependent wavenumber values,a short-space two-dimensional Fourier transform(FT)method is presented for processing wavefield data of Lamb waves.The results can be used to determine the location,size and depth of rectangular notch.The analysis techniques are demonstrated using simulation examples of an aluminium plate with a rectangular notch.Then,the wavenumber analysis method is applied to simulation data that are obtained through a range of notch depths and widths.The results are analyzed and rules of the technique with regards to estimating notch depth are determined.Based on simulation results,guidelines for using the technique are developed.Finally,experimental wavefield data are obtained in aluminium plates with rectangular notches by a full noncontact transceiving method,i.e.,laser-laser method.Band-pass filtering combined with continuous wavelet transform is used to extract a certain frequency component from the full laser-induced wavefield with wide band.Shortspace two-dimensional FT method is used for further processing full wavefield data at a certain frequency to estimate spatially dependent wavenumber values.The consistency of simulation and experimental results shows the effectiveness of proposed wavenumber method for quantitative rectangular notch detection.