光谱学与光谱分析, 2020, 40 (7): 2313, 网络出版: 2020-12-06  

深度自编码器的近红外光谱转移研究

Near Infrared Spectroscopy Transfer Based on Deep Autoencoder
作者单位
1 集美大学港口与环境工程学院, 福建 厦门 361021
2 温州大学机电工程学院, 浙江 温州 325035
3 温州大学电气与电子工程学院, 浙江 温州 325035
摘要
近红外光谱分析中多变量校准模型的建立依赖于校准建模的光谱样本。 然而, 近红外光谱测量环境的变化会导致同一被测物的光谱样本的偏移。 为了削减光谱偏移后重新建立校准模型的成本, 提出一种基于深度自编码器(DAE)的非线性光谱转移方法, 以端到端的形式实现不同测量环境之间的光谱转移, 避免已有的线性光谱转移方法在非线性偏移光谱时效果不佳的情况。 该方法在操作前不需要对光谱进行预处理和特征提取等操作, 可以实现原始光谱之间的转移, 是首个端到端的非线性光谱转移方法。 为了实现光谱空间的有效转移, 设计了一种基于条件概率和参数最大似然法的误差函数惩罚项, 结合梯度反向传播算法优化深度自编码的网络参数。 为了验证该方法的有效性, 引入两个公共的近红外光谱数据集, 分别是药片数据集和玉米数据集。 利用本方法进行光谱转移的过程主要有: 根据Kennard-Stone(KS)算法分别将两个数据集划分为校准集、 验证集和测试集; 用校准集中的光谱样本输入深度自编码器, 根据设计的误差函数求出误差, 并用反向传播法迭代训练网络参数, 直至模型最优; 将预测集样本输入训练好的DAE转移模型, 可以发现转移后的光谱与相应的目标光谱谱线基本重合, 这说明该设计的转移模型的有效性。 最后, 为了进一步验证本方法的优越性, 将该方法与经典的线性转移算法光谱空间变换(SST)和分段直接标准化(PDS)进行比较。 这三种算法得到的转移光谱分别作为测试样本, 输入已建立的偏最小二乘(PLS)多变量校准模型, 通过比较预测均方根误差(RMSEP), 可以发现该方法在多变量校准模型中的预测结果的均方根误差均小于SST和PDS, 分别提高了5.7%和10.1%, 表明由非线性深度自编码器转移的光谱样本具有高效和实用的特点。
Abstract
The establishment of a near-infrared spectroscopy multivariate calibration model relies on calibration samples. However, changes in the near-infrared spectroscopy measurement environment can cause differences between the spectra easily. In order to reduce the consumption of rebuilding calibration model on offset spectrum, this paper proposes a nonlinear spectral transfer method based on deep autoencoder (DAE), which can realize the spectral transfer in an end-to-end way. This method compensates for the poor performance of existing linear spectral transfer methods in the face of nonlinear offset spectra. Moreover, this method can realize the transfer between raw spectra without data process and feature extraction operations. In the paper, we propose an error function penalty term based on the conditional probability distribution and parameter maximum likelihood method, and combine it to the gradient back-propagation algorithm to optimize the parameters of DAE. In order to verify the effectiveness of the method, we perform the proposed method on tablet dataset and corn dataset, which are both public near-infrared spectral datasets. First, we divide the two datasets into the calibration set, validation set, and prediction set using Kennard-Stone (KS) method respectively. Then, we design a network structure that conforms to the spectral sample dimension of the dataset. Finally, samples in calibration set are input to the DAE, and the network parameters are iteratively optimized by the proposed error function and back-propagation algorithm. After the transfer model is established, we compare it with the spatial transformation (SST) and piecewise direct standardization (PDS), both of them are classical linear transfer algorithm. The transferred spectrum obtained by these three algorithms are respectively inputted into the established multivariate calibration model, and we can find that the root means square error of prediction (RMSEP) of the proposed method averagely improves 5.7% and 10.1% than SST and PDS respectively, which can be demonstrated that the spectral samples transferred by the nonlinear deep autoencoder are highly efficient and useful.

刘贞文, 徐玲杰, 陈孝敬. 深度自编码器的近红外光谱转移研究[J]. 光谱学与光谱分析, 2020, 40(7): 2313. LIU Zhen-wen, XU Ling-jie, CHEN Xiao-jing. Near Infrared Spectroscopy Transfer Based on Deep Autoencoder[J]. Spectroscopy and Spectral Analysis, 2020, 40(7): 2313.

关于本站 Cookie 的使用提示

中国光学期刊网使用基于 cookie 的技术来更好地为您提供各项服务,点击此处了解我们的隐私策略。 如您需继续使用本网站,请您授权我们使用本地 cookie 来保存部分信息。
全站搜索
您最值得信赖的光电行业旗舰网络服务平台!