Series fusion of scatter correction techniques coupled with deep convolution neural network as a promising approach for NIR modeling

Abstract

Deep convolution neural network (CNN) with one-dimensional (1D) convolution structure is a potential and effective nonlinear method for near infrared (NIR) spectroscopy analysis. However, it is also a challenge to build a reliable CNN calibration model since industrial NIR data present serious scattering effect which will seriously interfere with important information. Thus, this paper proposed a promising approach, namely series fusion of scatter correction technologies (SCSF), where CNN built on the series splicing data of normalized raw spectra, standard normal variable (SNV) spectra and first derivative (1d) spectra. Two real NIR cases (one is the identification of alcohols/diesel blends and the other is the prediction of methanol and ethanol content in alcohols/diesel blends) were introduced to explore the feasibility and effectiveness of the presented model. Through the comparative analysis with CNN based on raw spectra, SNV spectra and 1d spectra, as well as common support vector machine (SVM) and BP neural network, the proposed SCSF coupled with CNN cannot only achieve 97.73 % recognition rate for three types of diesel, but also significantly improve the prediction accuracy of methanol and ethanol. Satisfactory results show that SCSF approach can be regarded as series boosting of multiple scatter correction technologies to improve overall performance without mastering data prior information and professional knowledge. Further, the proposed SCSF applied to CNN deep learning is simple and efficient, and can be recommended for actual implementation in industrial NIR applications.