文献类型：期刊文献

英文题名：Multi-spectral fusion and self-attention mechanisms for Gentiana origin identification via near-infrared spectroscopy

作者：Li, Sihai[1];Wang, Yangyang[1];Song, Hang[1];Liu, Mingqi[1]

第一作者：李四海

通信作者：Li, SH[1]

机构：[1]Gansu Univ Chinese Med, Coll Pharm, Lanzhou 730000, Gansu, Peoples R China

第一机构：甘肃中医药大学药学院（西北中藏药协同创新中心办公室）

通信机构：[1]corresponding author), Gansu Univ Chinese Med, Coll Pharm, Lanzhou 730000, Gansu, Peoples R China.|[1073501e14fb35863569f]甘肃中医药大学药学院（西北中藏药协同创新中心办公室）;[10735]甘肃中医药大学;

年份：2024

卷号：246

外文期刊名：CHEMOMETRICS AND INTELLIGENT LABORATORY SYSTEMS

收录：;EI(收录号：20240615518534);Scopus(收录号：2-s2.0-85184020078);WOS:【SCI-EXPANDED(收录号:WOS:001178225100001)】；

基金：This study is financially supported by the National Natural Science Foundation of China (grant no. 82074419) , the Gansu Provincial Science and Technology Program Project of China (grant no. 21JR1RA272) , and the Gansu Provincial Department of Education: University Teachers Innovation Fund Project (grant no. 2023B-105) .

语种：英文

外文关键词：Gentiana; NIRS; Data Fusion; CNN; GRU

摘要：Gentiana is rich in Gentiopicroside and strychnine acid with medicinal value. However, the active ingredients of Gentiana from different origins are different, so identifying Gentian's origin is significant. Currently, neural networks such as CNN and GRU are widely used for spectral data analysis, but the modeling effect is easily affected by the spectral preprocessing method, and the long region and many features of spectral data make it difficult for CNN models to capture the long-term dependence of spectra, while GRU modeling has a large number of parameters, high computational complexity, and low efficiency. Therefore, a Gentian Root Data Fusion Module (GL) for sequence data is proposed to achieve the fusion between spectral data under different pre-processing by assigning weights to multiple pre-processing data and all features of pre-processing data respectively, making full use of the advantages of different pre-processing methods. Aiming at the characteristics of the long spectral data region, the joint architecture of convolutional neural network (CNN) and gated neural network (GRU) is adopted to achieve the extraction of features and the capture of long-term dependencies, while reducing the model complexity. Finally, GL is integrated with CNN and GRU to craft the advanced collaborative framework known as CCRN. The experimental findings demonstrate that CCRN outperforms CNN + GRU, CNN, PLS-DA, and SVM in terms of accuracy and loss function performance. Notably, CCRN exhibits superior Accuracy, Recall, and F1-score, surpassing the CNN + GRU model by 2.4 %, 2.1 %, and 2.1 %, respectively. These results validate the efficacy of the GL module in seamlessly integrating various preprocessing methods. In addition, the model CCRN still performs best when tested on public datasets, proving that CCRN has good Portability and scalability.