中国科学院自动化研究所机构知识库

   随着化工、食品、制药、医疗和农业等行业对产品质量改进和生产精细管控的要求越来越高，近红外光谱分析作为一种快速、非接触、成本低廉的测量技术逐渐取代传统化学实验室分析技术应用于工业质量控制和产品成分检测等领域中。近红外光谱分析技术成功应用的关键在于精确鲁棒的校正模型的构建。然而传统近红外光谱建模方法需要复杂的人工先验知识、精巧的预处理和变量选择方法。与传统近红外光谱建模方法相比，深度学习方法能够端到端地自动提取近红外光谱数据的抽象特征，无需复杂的预处理方法。本文主要研究了基于深度神经网络的近红外光谱校正模型的构建，同时探索了基于条件生成对抗网络的近红外光谱样本生成，并将模型应用于土壤近红外光谱数据集中有机碳含量的定量估测。主要研究内容和结论如下：

1.    基于深度神经网络的校正模型构建。本文提出一种结合卷积神经网络和 循环神经网络的模型架构称为 CCNVR 用于近红外光谱数据的建模。它主要结合了卷积神经网络能够直接从原始近红外光谱提取局部抽象特征和循环神经网 络能够学习到特征之间复杂依赖关系的特性。为了验证 CCNVR 模型的预测精确性和鲁棒性，将 CCNVR 模型与卷积神经网络、循环神经网络、人工神经网络、偏最小二乘回归和支持向量机回归模型的预测性能作比较，结果表明 CCNVR 模型具有更好的预测性能和鲁棒性。

2.    基于条件生成对抗网络的近红外光谱样本生成。本文提出一种改进的条件生成对抗网络模型 ACWGAN 用来生成带有目标成分含量值属性的近红外光谱样本，为了验证模型的有效性，将模型应用于不同十六烷值条件下柴油近红外光谱样本的生成，结果证明模型能够生成与真实样本高度相似但局部存在差异的近红外光谱样本，增加了样本的丰富度，提高了校正模型的预测准确性。

3.    基于深度学习的近红外光谱校正模型在土壤营养成分估测中的应用。本文将基于深度学习的近红外光谱校正模型应用到土壤中有机碳含量的估测。由于土壤样本分布极不均衡，深度学习方法在不平衡数据集上泛化能力较差。为了解决这一问题，本文先使用 ACWGAN 模型生成种类数目占比较少的近红外光谱样本，然后使用生成的样本扩充原有校正集，使得样本分布更均衡，再在扩充后的校正集上重新训练CCNVR 模型。结果表明使用 ACWGAN 生成的少类别样本扩充原有近红外光谱数据库后能够显著提高深度校正模型 CCNVR 在不平衡数据集上的预测准确性和泛化能力。

With the increasing demand for product quality improvement and fine control of production in the chemical, food, pharmaceutical, medical and agricultural fields, Near Infrared (NIR) spectroscopy is gradually replacing traditional chemical laboratory analytical techniques as a fast, non-contact, and cost-effective measurement technique for industrial quality control and detection of product components. The key to the successful application of NIR spectroscopy lies in the construction of accurate and robust calibration models. However, conventional calibration models require intricate manual prior knowledge, sophisticated preprocessing, and variable selection methods. In contrast to traditional calibration models, deep learning is an end-to-end method that can automatically extract abstract features from the raw spectrum without preprocessing. This paper focuses on constructing a calibration model based on deep neural networks and explores the generation of NIR spectral samples based on conditional generative adversarial networks，and then applies the model to the quantitative estimation of organic carbon content in soil NIR spectral datasets. The main research contents and conclusions are as follows：

1.  The construction of calibration models based on deep neural networks. To overcome these weaknesses, we proposed a jointed Convolutional Neural Network (CNN) and Recurrent Neural Network (RNN) architecture called CCNVR, which combines the ability of CNN to extract the local and abstract features from the raw spectrum with the advantage of RNN to learn various dependencies of sequence features. We then compared the prediction accuracy of CCNVR with other conventional methods, namely, partial least squares regression (PLSR), support vector machines regression(SVMR), CNN, artificial neural networks(ANN), and RNN, on the selected soil properties of mineral soil samples in the Land Use/Land Cover Area Frame Survey (LUCAS) database. Of all calibration models, our proposed CCNVR achieved the best prediction performance and robustness.

2.  The generation of NIR spectral samples based on conditional generation adversarial network. We proposed an improved conditional generation adversarial network model called ACWGAN to generate NIR spectral samples with target component content values. To verify our proposed model’s effectiveness, the ACWGAN model was applied to generate diesel NIR spectral samples with different cetane values. The results proved that the model could generate NIR spectral samples with a high degree of similarity to the real samples but with certain variations. Further, it increases the richness of the sample and improves the prediction accuracy of the calibration model.

3.    The application of a calibration model based on a deep learning model for soil nutrient estimation. In this paper, we applied a calibration model of NIR spectra based on deep learning to estimate organic carbon content in soils. Due to the highly uneven distribution of soil samples, the deep learning method has poor generalization ability on unbalanced data sets. To solve this problem, the ACWGAN model is used to generate NIR spectral samples with fewer species. Then the generated samples are used to expand the original calibration set to make the sample distribution more balanced. The CCNVR model is then retrained on the expanded calibration set. The results show that using the ACWGAN generated samples with fewer categories to expand the original NIR spectral database can significantly improve the prediction accuracy and generalization ability of the depth calibration model CCNVR on the unbalanced data set.