基于太赫兹波段的材料辨识技术研究

发布时间：2018-06-04 16:47

本文选题：太赫兹 + 太赫兹时域光谱　；参考：《国防科学技术大学》2015年硕士论文

【摘要】：太赫兹波是一种频率范围一般在0.1THz~10THz之间的电磁波,介于微波和红外之间。由于20世纪80年代之前由于缺乏大功率的太赫兹源和高灵敏度的探测器,这一波段一直没有得到有效的开发和利用,因此被称为太赫兹间隙(THz gap)。由于太赫兹波具有很多优良的性质,并且随着太赫兹源和探测器技术的提高,以太赫兹时域光谱技术为代表的太赫兹检测技术得到了广泛的应用。本文利用太赫兹时域光谱技术开展材料辨识技术研究,主要研究内容和成果如下:(1)针对传统的光谱平滑方法,参数无法自动设置,容易出现过平滑,使得光谱中的吸收峰被减弱甚至消除,同时具有无法去除脉冲噪声的缺点,提出了一种改进的分数阶积分光谱平滑方法,首先使用中值滤波器去除脉冲噪声,然后对分数阶积分平滑后的光谱进行加权平均,提高了算法对含多种噪声类型光谱的适应性,能够取得更好的平滑效果和更高的信噪比。(2)针对太赫兹光谱中存在的基线漂移现象,提出了一种非对称ε-不敏感支持向量机,通过对正的拟合误差和负的拟合误差使用不同程度的损失函数,使得拟合的曲线不再是针对光谱本身而是光谱基线,用原始光谱减去拟合的基线从而对光谱实现了基线校正,相对于传统方法,校正的结果可以更好的保持吸收峰的原始形状。(3)针对传统的太赫兹谱分类识别方法将太赫兹光谱单纯的作为一个普通一维向量,没有依据太赫兹光谱本身的特点来处理,导致识别率不高的现象,提出了一种基于多尺度分析的特征选择方法,能够有效的提取光谱特征吸收峰,去除噪声带来的干扰。同时针对支持向量机训练多类别、大规模数据时,内存消耗和时间消耗过大的缺点,提出了一种改进的DAG-SVM快速算法,该算法首先使用基于均值漂移的支持向量预选取方法,避免了支持向量漏选的现象,然后根据类别之间的邻接图对原始DAG-SVM算法的决策树进行修剪,能够减少训练器的个数,加快训练的速度。
[Abstract]:Terahertz wave is a kind of electromagnetic wave with frequency range between 0.1THz~10THz and infrared. Due to the lack of high power terahertz source and high sensitivity detector before 1980s, this band has not been developed and utilized effectively, so it is called THz gap. With the improvement of terahertz source and detector technology, terahertz detection technology, represented by terahertz time-domain spectroscopy, has been widely used. In this paper, terahertz time-domain spectroscopy is used to develop material identification technology. The main research contents and results are as follows: (1) aiming at the traditional spectral smoothing method, the parameters can not be set automatically, and it is easy to occur smoothing. The absorption peak in the spectrum is weakened or eliminated, and the pulse noise can not be removed. An improved fractional-order integral spectral smoothing method is proposed. First, the median filter is used to remove the pulse noise. Then, the weighted average of the spectrum after fractional integral smoothing is carried out, which improves the adaptability of the algorithm to the spectrum with multiple noise types. A novel asymmetric 蔚 -insensitive support vector machine is proposed to solve the baseline drift phenomenon in terahertz spectrum, which can achieve better smoothing effect and higher signal-to-noise ratio (SNR). By using different loss functions to the positive and negative fitting errors, the fitted curve is no longer aimed at the spectrum itself but the spectral baseline. The baseline correction is realized by subtracting the fitted baseline from the original spectrum. Compared with the traditional method, the corrected results can better maintain the original shape of the absorption peak.) the terahertz spectrum is simply regarded as a general one-dimensional vector for the traditional THz spectrum classification and recognition method. A new feature selection method based on multi-scale analysis is proposed, which can extract spectral characteristic absorption peak effectively and remove the interference caused by noise because it is not dealt with according to the characteristics of terahertz spectrum itself, which leads to low recognition rate. At the same time, an improved DAG-SVM fast algorithm is proposed to solve the problem of excessive memory and time consumption when SVM is training multi-class and large-scale data. Firstly, support vector pre-selection method based on mean shift is used in this algorithm. The missing support vector is avoided and the decision tree of the original DAG-SVM algorithm is pruned according to the adjacent graph between classes, which can reduce the number of trainers and speed up the training.
【学位授予单位】：国防科学技术大学
【学位级别】：硕士
【学位授予年份】：2015
【分类号】：TB30;O441

【参考文献】