Search

Abstract In infrared spectroscopy analysis area, quantitative modeling is often an essential and inevitable procedure to obtain the relationship between collected spectral information and target components. Over the past decades, there are many linear or nonlinear methods have been proposed, such as multi-regression, partial least squares, artificial neural network, support vector machine etc. However, these traditional methods commonly need some preprocessing steps including denoising, baseline correction, wavelength selection and so on. Hence, it requires the users to master so many skilled knowledges before they can establish a good performance quantitative model. Additionally, the stabilities of the above-mentioned methods are often not well enough because there are many random parameters which will result in the model's output are always changing every time. To solve these problems, this paper proposed an end-to-end modeling method based on convolutional neural network and ensemble learning. The experimental results on three infrared spectral datasets (corn, gasoline and mixed gases) showed that the generalized performance of proposed ECNN method outperforms traditional methods like PLS, BP neural network and single CNN method with whole spectral range raw data instead of selected wavelengths. Hence, the proposed method can obviously reduce the modeling knowledge for users and easier to use. Highlights • A novel end-to-end quantitative analysis modeling method for infrared spectroscopy was proposed based on convolutional neural network (CNN), which can directly take the whole range of collected raw spectral information as input without wavelength selectionpreprocessing. • An ensemble framework wasbuilt to improve the stability(also called “robustness”)of quantitative analysis model by introducing ensemble learning ideas. • The performance of proposed method is investigated through two public and one experimentaldatasets in the area ofinfrared spectroscopy. [ABSTRACT FROM AUTHOR]

Variable selection is an essential part during the whole process of qualitative and quantitative analysis of spectroscopy. Traditional methods like interval partial least square (iPLS), uninformative variable elimination (UVE), successive projections algorithm (SPA) etc. often have some disadvantages such as many parameters need to be tuned, weak robustness and so on. To solve these problems, this paper proposed a novel variable selection framework which combines UVE algorithm and ensemble L1 regularization framework together. The whole process of proposed method includes two phases: rough and fine selection. Firstly, UVE algorithm is used to eliminate the uninformative variables (rough selection). Secondly, the variable selection problem is mapped into a L1 regularization optimization problem with constraint (fine selection). To improve the stability and robustness of proposed method, an ensemble variable selection framework is designed which ensemble the results of many L1 regularization selectors. To validate the performance of proposed method, the following two public near infrared spectral datasets were tested: (1) Spectral (range from 900 nm to 1700 nm) and octane data of gasoline; (2) Spectral (range from 1100 nm to 2498 nm) and moisture data of corn. The experimental results showed that the proposed method can not only select the most featured wavelengths, but also can improve the stability and robustness of variable selection results. [ABSTRACT FROM AUTHOR]