Master thesis : Classification of aluminum alloy scraps via spectral analysis: A deep learning approach to laser-induced breakdown spectroscopy

Abstract

This work explores innovative methodologies in metal classification, focusing on aluminum, within the GeMMe laboratory’s framework and the SALEMA project. This research focuses on three main objectives: firstly, to maximize information extraction from the complete spectral data for aluminum classification; secondly, to deepen understanding of the set structure of the aluminum samples composed of multiple Laser-Induced Breakdown Spectroscopy spectra; and thirdly, to enhance classification techniques by integrating unsupervised datasets. This work employs various deep learning models, including the Multi-Layer Percep tron (MLP), Residual Network (ResNet), Variational Autoencoder (VAE), and their skip-connected variant. It compares them against GeMMe’s baseline in preprocessed and non-preprocessed data scenarios. Experimental findings suggest that a comprehensive analysis of the entire spectrum of each Laser-Induced Breakdown Spectroscopy spectra contributes to an improved weighted F1 of 2% compared to the GeMMe’s PICKIT baseline. Secondly, acknowl edging the set structure of the samples, further refines the classification, particularly with the application of deep set techniques. This approach has shown a significant increase in weighted F1 score of 11% compared to the GeMMe’s PICKIT baseline. Additionally, the introduction of unsupervised datasets in the context of generative models, particularly the skip-connected VAE, markedly enhanced the weighted F1 of 1.3%, surpassing the previously mentioned approach, especially for smaller sets. This was achieved by balancing the generative and the classification loss. The model’s generalization capabilities were tested on unseen datasets, showing weighted F1 over 85%, yet indicating challenges in samples with fewer spectra.