非均匀固体籽粒近红外光谱采集方法研究

近红外光谱定性分析技术研究中实验对象主要包括液态物质、粉末状物质和非均匀固体颗粒物质。液态物质与粉末状物质属于均匀分布被测对象，光谱采集与分析相对容易。而非均匀固体样品大小不一、形态各异、内部物质分布不均匀，近红外光谱中既包含定性分析中需要提取的种类信息，又包含需要消除的个体差异信息，因此相对于均匀分布的液体或粉末状物质，非均匀固体分析难度更大。目前国内外尚未建立对于大小不一、形状各异的非均匀固体颗粒物质的有效近红外定性分析方法。以非均匀固体谷物玉米籽粒为主要研究对象，在研究多种光谱采集方式的基础上，分析了非均匀固体籽粒光谱采集的特点，设计了针对非均匀固体籽粒的光谱采集装置，为确保实验结果的客观性，采用了朴素贝叶斯分类器、K近邻、支持向量机、BP神经网络、仿生模式识别等5种模式识别方法，建立在漫反射及漫透射方式下单籽粒玉米近红外光谱多品种定性分析模型，并对漫反射与漫透射方式下所建立的定性分析模型进行对比实验，分析胚面朝向对玉米单籽粒的漫反射与漫透射光谱定性分析模型鉴别效果影响，测试集与建模集时间间隔长短对漫反射及漫透射方式下玉米单籽粒品种鉴别精度的影响。实验结果表明采用漫透射方式建立的模型不易受到非均匀固体籽粒摆放方式的影响，且模型具有更好的泛化能力，为后续研究提供可行光谱采集方法。以非均匀固体玉米籽粒为主要实验对象，对其采集方法与定性分析模型进行研究，可为类似形态被测物体的近红外光谱定性分析提供有益借鉴，具有重要研究意义。

Study on Near-Infrared Spectrum Acquisition Method of Non-Uniform Solid Particles

In the research of near-infrared spectroscopy (NIRS) qualitative analysis technology, the experimental objects mainly include liquid matter, powder matter and non-uniform solid particles. The liquid and powder materials are evenly distributed, and the spectrum collection and analysis are relatively easy. The non-uniform solid particles are different in size, shape and internal material distribution. Near-infrared spectroscopy of these samples contain the type information to be extracted in qualitative analysis and the individual difference information to be eliminated. Therefore, the analysis of non-uniform solid particles is more difficult than that of liquid or powder materials with uniform distribution. At present, there is no effective NIRS qualitative analysis method for non-uniform solid particles with different sizes and shapes at home and abroad. In this paper, non-uniform solid grain maize is taken as the research object. Based on the study of various spectral acquisition methods, the characteristics of non-uniform solid grain spectral acquisition are analyzed, and the spectral acquisition device for non-uniform solid grain is designed. In order to ensure the objectivity of the experimental results, Five pattern recognition methods such as Naive Bayesian classifier, k-nearest neighbor, support vector machine, BP neural network and Biomimetic Pattern Recognition were used to establish the qualitative analysis model of single grain maize by near-infrared spectroscopy in diffuse reflection and diffuse transmission mode, and the qualitative analysis models established in diffuse reflection and diffuse transmission mode were compared to analyze the effect of embryo orientation on the identification of single grain maize. The effect of the time interval between train and test sets on the identification accuracy under diffuse reflection and diffuses transmission is studied too. The experimental results show that the diffuse transmission model is not easily affected by the lying style of non-uniform solid grains, and the model has better generalization ability, which provides a feasible spectral acquisition method for subsequent research. Taking the non-uniform solid maize grain as the main experimental object, the research on its collection method and qualitative analysis model can provide a useful reference for the qualitative analysis of near-infrared spectroscopy of similar objects, which has important research significance.