Abstract：Near-infrared spectroscopy(NIRS) technology is widely used in petroleum, textiles, food, pharmaceuticals, etc., due to its fast, non-destructive, and efficient characteristics. However, there are problems with traditional analysis methods,such as difficulty in feature extraction, low modeling accuracy when dealing with spectral data with many variables, and high redundancy. Therefore, this paper proposes a quantitative modeling method of one-dimensional wavelength attention convolutional neural network (WA-1DCNN) suitable for near-infrared spectroscopy without variable screening. The modeling method has a simple structure, strong versatility, and high accuracy.This study introduces the wavelength attention mechanism, which enhances the model's ability to capture important wavelength features by giving different weights to different wavelength data, thereby improving the accuracy and robustness of quantitative analysis. Four publicly available near-infrared spectral datasets were used in this paper to verify the feasibility of the proposed method. The proposed algorithm was compared with three traditional modeling methods that added wavelength screening, namely partial least squares (PLS), support vector regression (SVR), extreme learning machine (ELM), and one-dimensional convolutional neural network (1DCNN)modeling method. The model performance indicators root evaluated the model performance mean square error (RMSE) and coefficient of determination (R2). The results show that the performance indicators of the WA-1DCNN modeling method without the wavelength screening algorithm are better than those of the traditional modeling method and the 1DCNN modeling method with the wavelength screening algorithm. The R2 of the test set in the 655 tablets dataset is 0.956 3, which is 4.34%, 12.56%, 18.42%, and 11.59% higher than that of 1DCNN and PLS, SVR, and ELM with wavelength screening; the R2 of the test set in the 310 tablets dataset is 0.957 4, which is 2.72%, 8.28%, 7.27%, and 1.17% higher than that of 1DCNN and PLS, SVR, ELM, and 1DCNN with wavelength screening; The R2 of the test set were 0.980 3 and 0.968 5, respectively, which were 6.24%, 1.48%, 1.75%, 6.08% and 5.81%, 1.85%, 1.58%, 2.96% higher than those of 1DCNN and PLS, SVR, and ELM with wavelength screening; in the wheat protein dataset, the R2 of the test set was 0.960 0, which was 8.67%, 5.79%, 7.94%, and 0.56% higher than those of 1DCNN and PLS, SVR, and ELM with wavelength screening. To verify the optimality of the WA-1DCNN model structure in this paper, ablation experiments were conducted on four near-infrared spectral datasets to change the WA-1DCNN model structure. The results show that the wavelength-attention convolutional neural network is a spectral quantitative analysis method with strong versatility, high generalization ability and simple structure, which can promote the quantitative analysis of near-infrared spectra.
[1] Bec K B, Grabska J, Huck C W. Molecules, 2020, 25(12): 2948.
[2] LIU Yan, CAI Wen-sheng, SHAO Xue-guang(刘 言，蔡文生，邵学广). Science Bulletin(科学通报), 2015, 60(8): 704.
[3] Kamboj U, Guha P, Mishra S. Materials Today: Proceedings, 2022, 48: 576.
[4] Shen S, Hua J, Zhu H, et al. LWT-Food Science and Technology, 2022, 155: 112970.
[5] Ai N, Jiang Y, Omar S, et al. Molecules, 2022, 27(2): 335.
[6] Wu Y, Liu Y, Li X, et al. Fuel, 2022, 318: 123543.
[7] Devianti, Sufardi, Hafsah S, et al. Case Studies in Chemical and Environmental Engineering, 2023, 8: 100435.
[8] YU Shui, HUAN Ke-wei, LIU Xiao-xi, et al(于 水，宦克为，刘小溪，等). Spectroscopy and Spectral Analysis(光谱学与光谱分析), 2024, 44(6): 1627.
[9] YANG You, ZHOU Yu, LI Si-hai(杨 友，周 玉，李四海). Journal of the Chinese Cereals and Oils Association(中国粮油学报), 2024, 39(9): 198.
[10] Zhang M, Zhang T, Wang Y, et al. Chemosensors, 2024, 12(9): 184.
[11] Wang D, Zhao F, Wang R, et al. Frontiers in Plant Science, 2023, 14: 1138693.
