近红外光谱结合BP神经网络测定棉涤混纺面料的纤维含量

Detection of Fiber Contents of Cotton and Terylene Mixture Textile by Near Infrared Spectroscopy Combined with BP Neural Network

摘要: 通过近红外光谱结合误差反向传播的人工神经网络来检测棉涤混纺面料中纤维含量。测量了4000 cm–1~10,000 cm–1范围内棉涤混纺面料样品的近红外吸收光谱。利用小波变换滤波技术对吸收光谱数据进行压缩和去噪处理，结合滤波后重构光谱信号建立了棉涤混纺面料中棉和涤纶含量的BP神经网络校正模型。优化了隐含层神经元的节点数、学习率、动量因子和学习次数。对小波变换中的小波基和压缩尺度进行了详细的讨论。棉涤混纺样品的近红外光谱经过小波压缩，可以大大降低数据运算量。在小波尺度为3、隐含层神经元节点数为17时，模型的预测精度最高。所建立的棉和涤纶含量校正模型的预测集相关系数(RP)均为0.998，预测均方根误差为1.260%和1.860%。实验结果表明，应用傅里叶变换近红外光谱和BP神经网络技术来预测棉涤混纺面料纤维含量，可以满足定量分析的要求，该方法也适合于其他混纺面料纤维含量的快速测定。

Abstract: The prediction of fiber contents of mixture textile by near infrared spectroscopy (NIR) combined with back propagation (BP) neural network was investigated. The near infrared spectrum of samples with different cotton and terylene contents were obtained in the range of 4000 cm–1 - 10,000 cm–1. Wavelet transform (WT) was used for spectra data de-noise and compression. The correction model of cotton and terylene content based on BP neural network and reconstruction spectral signals was established. The number of hidden neurons, learning rate, momentum and epochs were optimized and decomposition levels of WT was discussed. Data procession was greatly reduced after the spectra signals were compressed by WT. When the compression level and the number of hidden neurons are 3 and 17 respec-tively, the prediction accuracy is the best. Correlation coefficients (RP) of prediction set for the correction model of cot-ton and terylene content both are 0.998, and the root-mean-square error (RMSE) is 1.260% and 1.860% correspondingly. Experimental results have shown that this approach by Fourier transform NIR based on the BP neural network to predict the cotton and terylene content of textile mixture can satisfy the requirement of quantitative analysis and is also suitable to other fiber contents measurement of textile mixture.