Classification of aluminum alloy using laser-induced breakdown spectroscopy combined with discriminative restricted Boltzmann machine

Abstract

Laser-Induced Breakdown Spectroscopy (LIBS), combined with modern machine learning tools, has emerged as a powerful technique for metal material identification, leveraging its high sensitivity and rapid response. However, the current spectral data analysis methods typically involve a two-step process of dimensionality reduction and model learning, lacking seamless integration. In this study, we address this issue by investigating a discriminative learning approach based on LIBS, utilizing the Discriminative Restricted Boltzmann Machine (DRBM). We apply LIBS technology in conjunction with DRBM for spectral feature selection and classification of five distinct small-sample aluminum alloy samples. The learned spectral latent distribution from the generative model component of DRBM effectively regularizes the discriminative process, thereby overcoming the problem of training overfitting arising from the high-dimensional small-sample limitation. This results in a stable and generalizable qualitative analysis model independent of empirical knowledge. The approach presented in this study achieves a 100 % accuracy, surpassing the best-performing traditional machine learning method (PCA-RF) by 13.33 % in accuracy and demonstrating a similar improvement compared to a Backpropagation Neural Network (BPNN) with the same structure.