近红外检测作为一种快速无损的检测方法得到广泛关注。但光谱中存在大量噪声以及光谱数据的高维度和非线性等特点影响了分类模型的准确率。将深信网络(DBN)的理论改进并引入光谱特征学习中, 解决高维特征间非线性关系的学习问题, 采用逐层训练策略和随机梯度上升法分别进行网络预训练和微调获得网络权值; 并结合支持向量机(SVM)建立近红外光谱多分类模型DBN-SVM。与基于主成分分析的分类模型PCA-SVM和基于线性判别分析的LDA-SVM分类模型进行应用比较。结果表明: DBN-SVM算法能有效地学习高维数据中的内在结构和非线性关系, 由该算法构建的模型具有良好的特征学习能力和分类识别能力, 而且在稳健性、各类别的灵敏度和特效度也更优。

As a fast and nondestructive detection method, near infrared detection has been widely concerned. But a lot of noises, high dimension and nonlinearity of the spectra affect the accuracy of classification model. In this study, deep belief network (DBN) theory was improved and introduced into spectral features learning to solve the difficulty of learning nonlinear relation of high dimensional data. The strategy based on layer by layer and stochastic gradient ascent algorithm were used for acquiring the network weights. Combined with the SVM method, the DBN-SVM multi classification model of tobacco leaf position was established. The proposed method, was compared with PCA-SVM method based on principal component analysis and LDA-SVM method based on linear discriminant analysis. The results show that DBN-SVM method may effectively learn the internal structure and nonlinear relationship of high dimensional data. The model constructed by this algorithm not only has excellent performance in identification and feature learning, but also is superior in robust, sensitivity and specificity.