近红外光谱属微弱信号, 其质量易受被测物体自身状态及各种外界因素干扰, 具体而言, 在近红外光谱定性分析中, 影响光谱质量的因素主要有光谱仪状态改变、 光谱采集人员错误操作、 奇异样本干扰等。 建模时若混入质量较差的光谱易影响所建模型的稳健性与适用性, 因此光谱质量判定是确保模型预测能力的一项重要工作。 目前用于定量分析的光谱质量判定研究较多, 而用于定性分析的光谱质量判定研究较少, 为此, 提出一种基于支持向量机数据描述的近红外光谱定性分析光谱质量判定方法, 采用自制漫透射近红外光谱装置采集单籽粒玉米光谱, 以正常状况下采集的某品种玉米单籽粒漫透射光谱作为正常样本, 而人为漏光、 近红外探测器窗口覆盖玉米表皮碎屑、 光源强度改变、 光源与被测玉米籽粒距离改变、 相近品种玉米籽粒混入等几种情况下所采集光谱作为异常样本, 在此数据集基础上研究了基于支持向量机数据描述的定性分析光谱质量判定模型建立的原理与方法, 其后将支持向量机数据描述方法与常用的马氏距离法、 局部异常因子法等光谱质量判定方法进行了对比, 并以正常样本正确识别率与异常样本正确拒识率的均值作为评价标准, 对实验结果进行分析, 由实验结果可以看出相比其他两种方法, 基于支持向量机数据描述的光谱质量判定方法具有最优判定能力, 建模集正常样本数目会影响光谱质量判定能力, 在实际使用光谱质量判定方法时, 建模集应包含足量样本。 在近红外定性分析时可以将该方法作为剔除异常光谱的手段, 在预处理、 特征提取, 模式分类等近红外光谱定性分析步骤前首先进行基于支持向量机的光谱质量判定步骤, 并剔除异常光谱, 可有效提高近红外光谱定性分析模型的可靠性, 亦为近红外光谱定性分析光谱质量判定提供新的方法参考。

Near infrared spectroscopy (NIR) is a weak signal, and its spectral quality is easily disturbed by the state of the measured object and various external factors. Specifically, the spectral quality in the qualitative analysis of NIR is mainly affected by the state change of measuring instrument, wrong operation, and the interference of singular samples. The robustness and applicability of the model are easily affected by the incorporation of poor quality spectra, so spectral quality determination is of vital importance to ensure the model prediction ability. At present, there are many studies on the determination of spectral quality for quantitative analysis, but few studies on the determination of spectral quality for qualitative analysis. In this paper, a method for the determination of spectral quality for near-infrared qualitative analysis based on data description of support vector is proposed. A self-made diffuse reflectance NIR acquisition device is used to collect the spectra of single-grain maize as an experimental object, and under normal conditions, the diffuse transmission spectra of a maize single grain were collected as normal samples, while the collected spectra were used as abnormal spectra under the conditions of artificial light leakage, near infrared detector window covering maize epidermis debris, intensity change of light source, distance change between light source and tested maize grain, and mixture of similar maize seeds. On this basis, the determination based on support vector data description (SVDD) was studied. The principle and method of establishing spectral quality judgment model were analyzed. Because the parameters of kernel function and regularization have important influence on the performance of spectral quality judgment model based on SVDD, the combination of grid search and cross validation was used to optimize the parameters of kernel function and regularization, and the optimal parameters of Gauss kernel were determined through experiments. Then, the SVDD method was compared with other spectral quality determination methods such as Mahalanobis distance and local anomaly factor. The average of correct recognition rate of normal samples and correct rejection rate of abnormal samples were used as evaluation criteria. The experimental results show that the spectral quality determination method based on support vector data description has the best performance. In near infrared qualitative analysis, this method can be used as a means of eliminating abnormal spectra before feature extraction and pattern classification, and the spectra quality determination step based on SVDD can effectively improve the reliability of the qualitative analysis.