Detection of Heavy Metal Content in White Peony by Laser-Induced Breakdown Spectroscopy Combined with Semi-Supervised Metric Learning
10.19756/j.issn.0253-3820.241416
- VernacularTitle:基于半监督度量学习的激光诱导击穿光谱检测白芍中的重金属含量
- Author:
Yan-Hong GU
1
;
Fan-Ding LI
;
Fu-Dong NIAN
Author Information
1. 合肥大学先进制造工程学院,合肥 230601
- Keywords:
Laser-induced breakdown spectroscopy;
White paeony;
Metric learning;
Consistency constraint;
Semi-supervised learning
- From:
Chinese Journal of Analytical Chemistry
2025;53(4):669-679
- CountryChina
- Language:Chinese
-
Abstract:
To address the economic challenge associated with acquiring labeled sample data for white paeony,a semi-supervised learning model based on metric learning and consistency constraints was proposed to predict the content of trace heavy metal pollutants Pb and Cd in white paeony.The model was comprised of two multi-task deep learning networks with the same structure but different parameters,namely the teacher model and the student model.The multi-task deep learning network utilized metric learning within the classification branch to ensure the clustering of different samples,thereby enhancing the predictive performance of the regression branch for heavy metal content in white peony samples.The student model effectively utilized unlabeled data by constraining the consistency of outputs between the teacher and student models.Experimental results showed that the proposed multi-task deep learning network combined with the regression subnetwork model significantly reduced the average relative errors of Pb and Cd in the test set to 7.01%and 8.16%when predicting trace heavy metal pollutants in paeony.Furthermore,after integrating the metric learning loss function-constrained and the consistency-constrained teacher-student semi-supervised learning model,the same samples exhibited clustering phenomena,with faster convergence speed and convergence values closer to 0 in the loss function,reducing the average relative errors of Pb and Cd in the test set to 3.32%and 4.77%.The above results indicated that the model proposed in this work could effectively enhance the accuracy and reliability of LIBS in quantitative analysis of trace heavy metal elements in paeony,strengthening the advantages of LIBS in practical applications for quality supervision in the traditional Chinese medicine market.
