小波多尺度正交校正在近红外牛奶成分测量中的应用

光谱分析中，干扰信号的存在直接影响所建分析模型的质量。基于信号和干扰的不同特性，提出了一种扣除背景和噪声干扰的新方法——小波多尺度正交校正(WMOSC)法。首先将原始光谱进行小波变换(DWT)，消除噪声及背景信息，然后采用正交信号校正(OSC)滤除与待测组分浓度无关的全部信息。与单纯的小波变换及正交信号校正相比，WMOSC能有效地扣除背景和噪声干扰，使模型具有更强的抗干扰能力，提高了模型的预测精度。利用该方法对牛奶样品的近红外光谱进行处理，采用偏最小二乘法建立校正模型，其牛奶中脂肪、蛋白质和乳糖的预测均方根误差(RMSEP)分别为0.101 6%，0.087 1%和0.110 7%。实验结果表明该方法能有效地去除干扰，保留有用信息。

Application of Wavelet Multi-Scale Orthogonal Signal Correction in Milk Components Measurement Using Near-Infrared Spectroscopy

Spectral interferences can have a significant impact on the spectral variation and as a consequence can adversely affect the results of calibration model in spectra analysis.Wavelet transform(WT)and orthogonal signal correction(OSC)were both the popular preprocessing algorithms.It was known that the former can effectively eliminate the background and noise and the latter can effectively filter out the interference information irrelevant to analyte concentration during the preprocessing of spectra.According to the different characteristics of analyte information and interference information in near-infrared(NIR)spectra,a new hybrid algorithm(WMOSC)that was the combination of discrete wavelet transform(DWT)and OSC was proposed to eliminate the spectral interferences including background,noise and systemic spectral variation irrelevant to the concentration.First,DWT was used to split the spectral signal into different frequency components,which keep the same data points as the original spectra data,to remove noise and background information by threshold method.Then OSC was applied to each frequency components to remove the information uncorrelated to the concentration independently.Finally,the spectra preprocessed by WMOSC were achieved through the summation of all frequency components.WMOSC was successfully applied to preprocess the NIR spectra data of milk.After elimination of the interference in the NIR spectra data by WMOSC,the partial least squares(PLS)regression was used to develop the calibration models for estimating the contents of main constituents in milk.The prediction ability and robustness of models obtained in subsequent PLS calibration using WMOSC were superior to those obtained using either DWT or OSC alone.The root mean square errors of prediction(RMSEP)of the models for fat,protein and lactose were 0.101 6%,0.087 1% and 0.110 7%,respectively.The experimental results show that WMOSC is an effective method for eliminating the interferences information in NIR spectra.