小波消噪及其在小麦蛋白质含量近红外光谱分析中的应用

WAVELET DENOISING THEORY AND ITS APPLICATION IN WHEAT PROTEIN CONCENTRATION PREDICTION WITH NEAR INFRARED SPECTROSCOPTY ANALYSIS

利用在小波变换下奇异信号和噪声在多尺度空间中的模极大值传递特性的不同,对33个小麦样品的近红外光谱信号进行了消噪处理,并利用小波消噪后重构光谱信号对小麦蛋白质含量进行偏最小二乘法交叉验证(PLS-CV)。计算实例表明,在最大分解层数不同时,PLS-CV效果各不相同,但大多数的小波消噪重构光谱进行PLS-CV,相关系数R及测定系数R2都有提高,交叉校验预测均方差RMSPCV都有减小,特别在最大分解层数为6时,PLS-CV效果最好,较使用原始光谱进行PLS-CV,相关系数R从0.9222提高到0.9698,交叉校验预测均方差RMSPCV从0.8014减小到0.4983。因此,使用小波消噪方法有消除原始光谱的噪声的作用,从而使最终的PLS模型更有代表性、稳定、稳健,也提高了品质检测时模型预测精度。

Based on wavelet denoising by using the difference in wavelet modulus maxima evolution behaviors between singular signals and random noises in multi-scale space, the NIR (near infrared spectroscopy) signals of 33 wheat samples were denoised and some PSL-LOO-CV (partial least squares-leave one out-cross validation) operations were proposed for the prediction of wheat protein concentration with the reconstructed spectra after denoising. The calculated instance showed that the PSL-LOO-CV results were not the same when maximum wavelet decomposing level was different, but both R (correlation coefficient) and R^2 (coefficient of determination) were improved on the whole and RMSPCVs (root mean squared errors of prediction of cross validation) were all reduced. PSL-LOO-CV results were best at a maximum wavelet decomposing level of 6, its R increasing from (0.9222 )to 0.9698 and its RMSPCV decreasing from 0.7880 to 0.4983 with the reconstructed spectra replacing the original spectra. It is, therefore, concluded that wavelet denoising is a useful method to eliminate noise of NIR signals for it makes the final PLS model more representative and more stable and robust, thus improving the prediction accuracy of quality detection.