小波包熵的复杂体系近红外光谱信息提取

有用信息提取是复杂体系近红外检测的重点和难点之一。由于复杂体系光谱中存在各种噪声、基线漂移、谱带重叠及复杂背景的干扰，常规方法不能准确地从光谱中获得有用信息。为此，将小波包变换(DWPT)和信息熵理论相结合--小波包熵(EWPIE)提取复杂体系光谱中的有用信息。思路是采用小波包变换对光谱信号进行多频带分解，根据有用信号与噪声的频带分布特点，基于信息熵理论滤除干扰的频率分量，采用正交校正法(OSC)剔除与被测组分无关的信息，然后对处理后的频率分量进行重构，从而实现复杂体系有用信息的准确提取。通过对复杂体系光谱数据建立多元校正模型来验证该方法的效果。采用牛奶的近红外光谱数据，以牛奶中脂肪和蛋白质浓度为研究对象，建立了偏最小二乘法(PLS)模型。结果显示，牛奶中脂肪和蛋白质的预测均方根误差(RMSEP)分别为0.132%和0.121%，与单纯的DWPT和OSC相比，EWPIE能够有效地提取有用信息，避免了无用信息的干扰，明显提高了模型的预测精度，对复杂体系的准确检测具有一定的理论意义和实际应用价值。

Information Extraction of Near Infrared Spectra for Complex Samples Based on Wavelet Packet Transform and Entropy Theory

The method of useful information extraction is one of the most important one in the detection application for complex samples using near infrared (NIR)spectroscopy.Due to various noise,baseline drift,peak overlapping and complex back-ground,informative spectra were hidden and cannot be extracted by conventional methods.Thus,a new hybrid algorithm (EW-PIE)was proposed for informative spectra extraction based on wavelet packet transform (WPT)and entropy theory in this stud-y.In EWPIE algorithm,WPT algorithm and its reconstruction algorithm were adopted to split the raw spectra into serval sub-spectra in different frequency bands.To take advantage of the multiscale property of NIR spectroscopy,each subspectra was fur-ther processed through entropy-based phase and orthogonal signal correction (OSC)-based phase.In entropy-based phase,the information entropy theory was used as a filter to remove the interference,which are uncorrelated to analyte contents and would increase the uncertainty of whole spectra.According to the variation of entropy value,some subspectra representing basedline in low frequency band and some subspectra representing noise in high frequency band were filtered out.In OSC-based phase,the OSC algorithm was applied to each of the remaining subspectra and the useful spectra were obtained with accumulation of all the OSC-filtered subspectra.To validate this algorithm,a real NIR spectral dataset of milk was prepared to extract the correlated spectra about the content of fat and protein.Using the EWPIE-filtered spectra,a series of PLS prediction model were construc-ted.Experimental results show that the prediction ability and robustness of PLS_based prediction models developed with EWPIE algorithm are superior to those developed by conventional algorithms.The root mean square errors of the prediction models for fat and protein can reach up to 0.132％ and 0.121％,which indicates that the EWPIE algorithm is a promise tool to extract the useful information from NIR spectra and has certain theoretical significance and practical application value for detection in com-plex systems.