Qualitative and Quantitative Analysis of Metal Scraps based on Transfer Learning of Certified Reference Materials Measurement in Laser-induced Breakdown Spectroscopy

Abstract

Laser-induced breakdown spectroscopy (LIBS) has been effectively used for the identification and quantification of the elemental composition of alloys but is challenged by its dependence on underlying experimental and environmental conditions. Emission spectra measured from the samples of the same type can vary significantly depending on the experimental setup such as the wavelength of the laser used, the detection range and spectral resolution of the spectrometer, the reflectivity of the collection mirror, and the movement of the sample. Besides, surface contamination, temperature, humidity, the time interval between experiments and arbitrary shape and size of the sample, etc., can cause extra variations in the spectral dataset. The framework of all pre-existing chemometric models is limited by the fact that both training and testing datasets should be obtained under identical or at least similar circumstances, which may not be practically feasible in many industrial applications. Moreover, variability in the metal scrap samples and a smaller number of measurements representing a class of samples add to the inadequacy of the training set of machine learning models. Therefore, this dissertation focuses on developing models that utilize certified reference materials (CRMs) datasets to classify and quantify the composition of metal scraps based on LIBS spectra, measured under varying experimental configurations and environmental factors. This dissertation largely consists of two parts; 1) data augmentation and transfer learning for qualitative analysis of metals, and 2) unsupervised domain adaptation for quantitative analysis of metal scraps.

In the first part of this dissertation, to improve the accuracy of the machine learning-based classification model, we proposed augmentation of the CRMs dataset by synthesizing spectrum through spectral line attenuation and generating spectra with a generative adversarial network (GAN) from available types. Thereafter, the augmented CRM dataset regarded as a source model was used to transfer knowledge through asymmetric heterogeneous transfer learning. This model was further fine-tuned according to the incoming testing dataset specification (called test-assisted tuning). Through analysis and simulation results, it was observed that both feature space and feature distribution varied in all the collected datasets, and final classification accuracy was able to outperform the results recorded for models based on prior information.

In the second part of the thesis, we proposed a transfer learning-based approach that depends on the domain adaptation of a source domain to transfer knowledge and reduce the difference between domains and thus improving the performance of the target learner. A source model pretrained with the standard (CRM) dataset is adapted in an unsupervised manner for the (metal scraps) test/target dataset based on a discriminator and adversarial loss for the unlabeled target datasets. For unsupervised domain adaptation and testing, metal scraps from actual industrial fields are tested using LIBS under different conditions such as lasers of different wavelengths and moving stages. The experimental results demonstrate that the proposed scheme achieves an impressive result for all different test datasets, which is significantly better than the results of conventional regression schemes. The proposed quantitative analysis approach can be generalized to the regression of metal scraps and other industrial applications using LIBS.

In summary, this dissertation comprehensively presents transfer learning-based machine learning models for qualitative and quantitative analysis of metal scraps for efficient recycling based on the LIBS measurement obtained from CRM samples.