Research on Two-Dimensional Convolutional Neural Network Model for Near Infrared Spectroscopy Analysis Based on Competitive Adaptive Reweighted Sampling and Gramian Angular Difference Field
10.19756/j.issn.0253-3820.251109
- VernacularTitle:基于竞争性自适应重加权采样和格拉姆角差场的二维卷积神经网络近红外光谱分析模型研究
- Author:
Xiao-Song ZENG
1
;
Ke-Wei HUAN
;
Xiao-Xi LIU
;
Xian-Wen CAO
;
Xue-Yan HAN
Author Information
1. 长春理工大学物理学院,长春 130022
- Keywords:
Near infrared spectroscopy;
Variable selection;
Gramian angular difference field;
Convolution neural network;
Quantitative analysis
- From:
Chinese Journal of Analytical Chemistry
2025;53(6):955-966
- CountryChina
- Language:Chinese
-
Abstract:
Near infrared spectroscopy(NIRS)analysis technology has become an important process analysis tool in industrial and agricultural production,and has been widely used for qualitative and quantitative analysis in the fields of tobacco,agriculture,and pharmaceuticals.To address issues such as poor generalization ability and low prediction accuracy in NIRS modeling,a two-dimensional convolutional neural network(2DCNN)quantitative analysis model based on competitive adaptive reweighted sampling(CARS)and Gramian angular difference field(GADF)(CARS-GADF-2DCNN)was proposed.CARS-GADF-2DCNN used the CARS method to select an optimal wavelength set from the full spectrum,then employed GADF to encode the selection results into two-dimensional images,and finally used 2DCNN for prediction analysis.The 2DCNN model consisted of convolutional layers,parallel convolution modules,flattening layer,and fully connected layers.Simulation experiments were conducted on three public near-infrared(NIR)spectral datasets encompassing soil,tablet,and grain datasets to evaluate the CARS-GADF-2DCNN model.The results demonstrated that,compared to the one-dimensional convolutional neural network(1DCNN),the GADF-2DCNN model achieved 16.74％,23.40％,and 7.13％improvement in prediction accuracy for the soil,tablet,and grain datasets,respectively.Compared to GADF-2DCNN,VCPA-GADF-2DCNN,and IRIV-GADF-2DCNN models,the CARS-GADF-2DCNN model further improved prediction accuracy.For the soil dataset,prediction accuracy improved by 39.00％,30.78％and 4.13％;for the tablet dataset,the improvements were 9.52％,6.94％and 2.56％;for the grain dataset,the improvements were 20.57％,9.85％and 15.66％.In conclusion,CARS-GADF-2DCNN effectively selected the optimal wavelength subset from near infrared spectra,and revealed the latent features between different wavelengths.CARS-GADF-2DCNN addresses the issues of high complexity in prediction models and low prediction accuracy in near infrared spectral modeling,and could be effectively applied to near infrared spectral prediction analysis of different substances.
