小波阈值降噪模型在红外光谱信号处理中的应用研究

摘要

Aimed at noise interference of infrared spectra, an example of using infrared spectra to detect fat content value on the surface of cashmere was applied to evaluate the effect of wavelet threshold denoising. The denoising capabilities of three wavelet threshold denoising models (penalty threshold denoising model, Brige-Massart threshold denoising model and default threshold denoising model) were compared and analyzed. Denoised spectra and measured cashmere fat content values were used for calibra-tion and validation with multivariate analysis (partial least squares combined with support vector machine). The authors analyzed and evaluated denoising effects of these three wavelet threshold denoising models by comparing parameters (R~2 , RMSEC and RMSEP) obtained through calibration and validation of denoised spectra with these three wavelet threshold denoising models re-spectively. The results show that the three wavelet threshold denoising models all can denoise the infrared spectral signal, in-crease signal to noise ratio and improve precision of prediction model to some extent; Among these three wavelet threshold de-noising models, the denoising effect of Brige-Massart threshold denoising model and default threshold denoising model were sig-nificantly better than that of default threshold denoising model; Compared with the prediction precision(R~2 =0.793, RMSEC= 0.233, RMSEP=0.225)of multivariate analysis model established with original spectra, the prediction precision(R~2= 0.882, RMSEC=0.144, RMSEP=0.136)of multivariate analysis model established with spectra denoised by Brige-Massart threshold denoising model and the prediction precision (R~2= 0.876, RMSEC= 0.151, RMSEP= 0.142)both had much more improve-ments. All the above illustrates that wavelet threshold denoising models can denoise infrared spectral signal effectively, make multivariate analysis model of spectral data and measured cashmere fat values more representative and robust, and so it can im-prove detection precision of infrared spectral technique.%针对近红外光谱经常受到噪声干扰的特点,提出了利用小波阈值降噪方法进行光谱数据的降噪处理,以山羊绒表面油脂的近红外光谱检测为例,对比分析了三种小波阈值降噪模型(Penalty阈值降噪模型、Brige-Massart阈值降噪模型、缺省阈值降噪模型)的降噪性能.对降噪后的光谱数据采用偏最小二乘和支持向量机回归相结介建立了校正和预测模型,通过对比校验参数R~2,RMSEC,RMSEP,分析评价了三种小波阈值降噪模型的降噪效果.结果表明:三种降噪模型都能在一定程度上降低光谱信号的噪声,提高信噪比,改善光谱预测模型的精度,其中,Brige-Massart阈值降噪模型和缺省阈值降噪模型的降噪效果明显优于Penalty阈值降噪模型,与原始光谱信号建模的预测精度(R~2=0.793,RMSEC=0.233,RMSEP=0.225)相比较,经过Brige-Massart阈值降噪模型降噪后的光谱信号建模的预测精度(R~2=0.882,RMSEC=0.144,RMSEP=0.136)和经过缺省阈值降噪模型降噪后的光谱信号建模的预测精度(R~2=0.876,RMSEC=0.151,RMSEP=0.142)均有较大程度的改善和提高,说明提出的小波阈值降噪方法能有效地降低原始光谱噪声作用,使光谱数据多变量分析模型更具有代表性和稳健性,从而可以提高模型的预测精度.