中国科学院半导体研究所机构知识库

近红外光谱分析技术具有便捷、无损、无化学试剂污染等优点，已在食品、药品、农业、石油石化、烟草、生物医学、纺织等领域获得了广泛应用。非均匀固体籽粒大小不一、形状各异，其光谱既包含类别信息，又包含待消除的个体差异信息，因此，定性分析研究中，非均匀固体籽粒的采集与分析存在较大难度。此外，目前国内外关于近红外光谱定性分析方法的研究偏向技术应用层面，缺乏系统性。同时，近红外光谱属于微弱信号，易受样品自身及各种外界因素干扰，导致定性分析模型存在鲁棒性欠佳问题。上述问题阻碍了近红外光谱定性分析技术的推广应用。鉴于此，本文以非均匀固体籽粒形态的玉米种子及小麦种子为研究对象，从以下方面对近红外光谱定性分析方法进行了研究，并取得创新性成果。

（1）针对大小不一、形状各异的非均匀固体被测对象光谱采集与分析问题，首先分析了非均匀固体光谱采集与分析特点，在此基础上，设计了漫反射与漫透射采集装置，采用5种模式分类方法建立非均匀固体漫反射光谱定性分析模型与漫透射光谱定性分析模型。以多个品种玉米籽粒为研究对象，通过实验对比分析两种光谱采集方式下模型性能差异。实验结果表明，漫透射方式比漫反射方式更适合于非均匀固体玉米籽粒近红外多品种鉴别，在漫透射方式下，多品种单籽粒玉米平均识别率可达90%以上。

（2）针对近红外光谱定性分析模型性能易受光谱质量影响的问题，提出了一种基于支持向量机数据描述（SVDD）的光谱质量判定方法。以玉米单籽粒漫透射光谱数据为例，通过向正常光谱中掺杂实际可能出现的异常光谱，将该方法与其他光谱质量判定方法进行了对比实验研究。实验结果表明，基于SVDD的光谱质量判定方法对多种异常光谱的拒识率能够达到90%以上。

（3）以玉米单倍体与二倍体近红外光谱为研究对象，研究了基于遗传算法的特征波长点选择方法。首先采用遗传算法与支持向量机分类器结合的方法从原始光谱中提取最有利于单倍体与二倍体分类的10个特征波长点，再利用10个特征波长点吸光度，采用支持向量机分类器对单倍体与二倍体进行分类鉴别。实验结果表明，使用上述10个特征波长点吸光度可以达到全光谱鉴别的效果。

（4）针对近红外定性分析模型的参数组合优化问题，提出了一种基于粒子群算法的定性分析模型参数组合寻优方法。采用粒子群算法对定性分析模型中平滑系数、一阶导系数、偏最小二乘特征提取维数、正交线性判别分析特征提取维数等参数进行寻优。实验结果表明，粒子群算法能够高效地搜寻到满足模型最优性能的特征提取参数组合，且代入最优参数组合的定性分析模型具有较好的泛化能力。

（5）针对种衣剂对玉米品种真实性鉴别准确性影响的问题，提出了一种基于栈式自编码神经网络（SAE）的近红外光谱定性建模方法。首先采用无种衣剂玉米籽粒光谱作为训练集，通过栈式自编码无监督学习算法与softmax分类器构建栈式自编码网络定性分析模型，再利用所建模型对有种衣剂玉米籽粒进行品种真实性鉴别。实验结果表明，基于SAE的建模方法能够将种衣剂对玉米籽粒识别率的影响降低至3%以内。

（6）针对单倍体玉米籽粒快速分选问题，提出了一种基于深度信念网络（DBN）的近红外光谱定性建模方法。首先采用多层受限玻尔兹曼机无监督学习及BP神经网络有监督学习结合训练得到多品种单倍体二倍体籽粒光谱的深度信念网络模型，然后对各个品种的单倍体二倍体籽粒进行分类识别。实验结果表明，使用DBN方法建立的多品种单倍体鉴别模型具有较强的分类性能，实验中所用10个品种的单倍体识别率均能达到90%以上。

（7）针对近红外定性鉴别模型鲁棒性问题，提出了利用样本多样性的模型鲁棒性增强方法，研究了基于多天联合建模的鲁棒性增强方法。首先使用多样性光谱数据与建模光谱数据联合构建偏最小二乘（PLS）空间，其后将建模光谱数据向PLS空间投影，再进行下一级特征提取与分类鉴别。实验结果表明该方法可以有效增强近红外定性鉴别模型鲁棒性，在模型参数不变的情况下，增加建模集所含品种数目时，模型识别率始终能保持在90%以上。在多天联合建模方法中，以逐日递增方式扩大联合建模集规模，再对多日测试集数据进行测试。实验结果表明多天联合建模方法可以有效延长模型适用期限，利用连续7日采集的单日数据联合建模，所得定性分析模型的识别率能够连续15日高于90%。

With the advantages of fast and nondestructive and pollution-free, near infrared spectroscopy (NIR) analysis technology has been widely used in the fields of food, pharmacy,agriculture, petroleum, petrochemical, tobacco, biomedicine, textile and so on. At present, more efforts mainly focus on NIR spectroscopy quantitative analysis researches, but few researches concentrated on NIR qualitative analysis method. Researchers mainly concentrated on application of NIR qualitative analysis in worldwide. In addition, solid irregular object such as corn seeds with different sizes and shapes is not studied in-depth, the spectrum of irregular object contains category information, and it also contains individual information need to be eliminated.so the qualitative analysis of irregular solid grain compared with ordinary uniform sample is more difficult, the research of NIR spectroscopy of non-uniform solid grain object is not researched abundantly. The shortage of poor robustness for qualitative analysis is also not researched in-depth, which have hindered the popularization and application of NIR qualitative analysis technique. In order to solve above problems, following key problems of NIR spectroscopy qualitative analysis methods are researched in this paper by taking the maize seeds and wheat seeds on solid and irregular state as the research object .

In order to solving the problem of multi-species identification for irregular particles with different sizes and shapes, the difficulties of spectral collection and analysis of irregular solid particle are analyzed firstly in this thesis The NIR diffuse reflectance model and diffuse transmission model for corn seeds identification were established by using several classification methods, and the performance differences of the two spectral acquisition methods were compared and analyzed by designed experiments. The experiments showed that diffuse transmission is more suitable than the diffuse reflectance in NIR identification of corn seeds and the recognition rate can reach higher than 90% with diffuse transmission method.

A SVDD based spectroscopy quality detection method of the NIR qualitative analysis is put forward in the thesis. A comparative study of several spectroscopy quality detection method was carried out by doping the artificial anomalies spectra in the normal spectra under the mode of diffuse transmission of maize single seed. The experimental results show the SVDD method can eliminate the abnormal spectroscopy effectively.The correct rejection rate of abnormal samples is higher than 90%.

From the aspects of reducing instrument cost and manufacturing difficulty, and aiming at haploid identification task, genetic algorithm (GA) to select characteristic wave points is studied. It is proved that the absorbance of 10 characteristic wave points can achieve full spectroscopy identification effect in the maize haploid identification task by the experimental results.

Aiming at the problem of combinatorial optimization problem of feature extraction, a method based on particle swarm optimization algorithm (PSO) is was proposed. The optimal parameter combination of preprocessing and feature extraction in the qualitative analysis model was designed to finding by using particle swarm algorithm. Then, the model was tested by using substitution optimal parameter combination into the qualitative analysis model. The results not only show that the PSO method can search the combination of preprocessing and feature extraction parameters which satisfy the optimal performance of the model efficiently, but also the qualitative analysis model with the optimal parameter combination has better generalization ability.

A method based on Stack Auto Encoder(SAE) neural network is was proposed. The corn spectrum without coat was used as training set, and the network SAE analysis model was established by using the Stack Auto-Encoding which is unsupervised learning algorithm and Softmax classifier. Then, the corn seed with coat was classified and identified by using the established model. The experimental results indicates SAE-based method can effectively reduce the influence of seed coat, the influence of seed coat can be reduced under 3% with SAE-based method.

The haploid discrimination method based on Deep belief network (DBN) was put forward aiming at the problem of rapid sorting of haploid. Firstly, a multilayer belief network model of multiple haploid grain spectra was obtained by using multilayer Restricted Boltzmann Machine and BP neural network method. Then, the haploid and diploid of each corn species were classified. It was proved that the multiple species haploid identification model by using DBN method has higher classification performance, and the haploid identification rate of 10 varieties can be higher than 90% in the experiment.

Aiming at the problem of poor robustness of NIR qualitative identification model, methods based on sample diversity and multi-day joint modeling are studied in this paper. Firstly, the partial least squares (PLS) space is constructed based on the method of sample diversity by using the diverse history data and modeling data. Then, the modeling data was projected into the PLS space, and the feature was extracted and classified. The experimental results show that the proposed method can improve the robustness of model effectively, and recognition rate can be kept at more than 90% while increasing the number of varieties with model parameters unchanged. Besides, multi-day joint modeling set was expanded by increasing the number of single-day data firstly, and then the multi-day Joint model was established, and model performance is tested with test data of several days. The experimental results show that the recognition rate can be higher than 90% for at least 15 days with seven-day joint model.

The object of this paper is irregular solid particles from the agricultural field with a variety of shapes and sizes, while the method and the conclusions proposed in this paper can be extended to other areas and other forms of object.