基于DS算法的玉米近红外定性分析光谱校正方法研究

Study on Spectral Calibration of Discrimination of Corn Variety Using Near-Infrared Spectra Based on DS Algorithm

从校正的角度出发,研究了近红外定性分析中模型稳定性问题。以13个玉米品种为研究对象,针对数据采集时间不同带来的模型失效问题,借鉴近红外光谱定量分析中两台仪器间模型传递的思想,将直接模型传递(Direct Standardization)算法用于校正同一仪器不同时间采集的光谱,使得一次建立的品种鉴别模型,能用于其余时间测试数据的鉴别。首先采用Kennard/Stone算法在主光谱集中选取校正样品集,按照对应的编号从从光谱集中取出对应的数据,然后对校正样品集采用DS算法求取两组数据间的变换关系,再对剩余的从光谱集进行相应的变换得到适用于模型的光谱。实验中对比了校正样本数和模型校正位置对校正结果的影响,分别从品种定性鉴别准确性和校正前后主光谱数据和从光谱数据分布距离两方面分析了实验结果。结果表明,该方法能有效地解决同一仪器随着采样时间推移产生的光谱偏移现象,对采样时间不同的测试集均得到较高的识别率,提高了模型的鲁棒性和适用范围,由实验结果可见,校正位置处于特征提取之后时,校正效果最佳。

From the perspective of calibration,the present paper studies the model stability problem in qualitative analysis of NIR.Aiming at the issue of model failure caused by different data acquisition time,13 varieties of corn were used as experimen-tal material,and learning from the idea of model calibration transfer between the two instruments in quantitative analysis of NIR,the DS（direct standardization ）algorithm was used to calibrate the spectra acquired at different times with the same instru-ment,that made the varieties identification model established one time able to be applied to identify the test data at different ac-quisition time.First,transfer set was selected from the master spectrum set by Kennard/Stone algorithm,the corresponding number spectrums in slave spectrum set were selected,and then DS algorithm was applied to transfer set to calculate the trans-formation function between the two sets of data.Finally,the remaining slave spectrums were transformed so that they could ap-ply to the model.This study does some experiment to discuss the impact of the number of transfer set and the location of calibra-tion on the calibration results.Respectively,the experiment results were analyzed from two aspects,one is the correct discrimi-nation rate in qualitative analysis,and the other is the distribution distance between master spectrums and slave spectrums before and after calibration.The experiment results indicate that this approach is effective to solve the spectra drift produced by sam-pling over time,can bring higher recognition rate on different sampling time test sets,also improves the robustness and applica-tion scope of the identification model,and the experiment results also indicate that the best result can be obtained with calibration locating after feature extraction.