A transferred multitask regularization convolutional neural network (TrMR-CNN) for laser-induced breakdown spectroscopy quantitative analysis. Issue 10 (13th September 2022)

• Record Type: Journal Article
• Authors: Cui, Jiacheng; Song, Weiran; Hou, Zongyu; Gu, Weilun; Wang, Zhe
• Is Part Of: Journal of analytical atomic spectrometry. Volume 37:Issue 10(2022)
• Page Start: 2059
• Page End: 2068
• Publication Date: 2022-09-13
• Subjects: Atomic spectra -- Periodicals; Atomic absorption spectroscopy -- Periodicals; 543.0858
• DOI: 10.1039/d2ja00182a
• Languages: English
• ISSNs: 0267-9477

Abstract

Abstract : A quantification method combining transfer learning, a convolutional neural network and multitask regularization to improve prediction accuracy and model robustness on limited data. Abstract : Laser-induced breakdown spectroscopy (LIBS) combined with machine learning has demonstrated great capabilities for quantitative elemental analysis. When the distributions of training and test data differ due to changes in measurement and sample composition, machine learning models degrade in accuracy and reliability. This, coupled with the small sample problem caused by high cost and long time to certify the analyte content, poses a challenge to the performance of LIBS quantification. This work proposes a transfer learning method to improve limited sample size LIBS quantification performance with extra-spectrum from similar LIBS measurements out of the experimental series. The model inherits convolutional layers from a source neural network model and is trained on a small amount of target training samples. Multitask regularization is introduced to constrain the source model based on prior information of the sample composition. The experiments were designed based on coal datasets with a limited sample size and different analyte concentration ranges. The proposed method reduces RMSEp by 19.9%, 5.9% and 7.7% compared to PLSR, SVR and non-transfer CNN models. The results show that the proposed method can outperform baseline approaches in terms of accuracy and robustness on limited data sets and can take effect on various tasks.