基于深度置信网络的红外光谱鉴别汽油掺混

为实现掺混汽油快速无损鉴别，提出一种利用t 分布邻域嵌入结合深度置信网络的鉴别方法，以解决机器学习中高维特征向量间的非线性关系。以92#、95#、98#及定比混合汽油为研究对象，采用多元散射校正算法对原始红外波段投射光谱测量数据进行预处理，利用t-SNE非线性方法进行光谱数据降维处理，分别采用深度置信网络和极限学习机建立汽油种类光谱鉴别模型并对比分析两种方法识别精度。研究表明:该文所选择方法构建的汽油鉴别模型性能更优，对汽油种类预测精准度高达92.5%，从而验证了该方法在汽油鉴别中的有效性。研究结果可为掺混成品油鉴别及溯源研究提供技术支持。

Identification of gasoline blending by infrared spectroscopy based on deep belief networks

In order to realize the fast nondestructive identification of blended gasoline, an identification method based on t-distributed stochastic neighborhood embedding(t-SNT) combined with deep belief networks was proposed to solve the nonlinear relationship between high-dimensional feature vectors in machine learning. Taking 92#, 95#, 98# and fixed ratio blended gasoline as the research objects, the projection spectrum measurement data in original infrared band was preprocessed by multivariate scattering correction algorithm, and the dimension reduction of spectral data was carried out by using t-SNE nonlinear method. The spectral identification model of gasoline types was established by using deep belief networks and extreme learning machine respectively, and the identification accuracy of the two methods was compared and analyzed. The research shows that the gasoline identification model constructed by this method has better performance, and the prediction accuracy of gasoline types is as high as 92.5%, which verifies the effectiveness of this method in gasoline identification. The results of this research can provide technical support for the identification and traceability of blending refined oil products.