Few-shot classification of coal and gangue by laser-induced breakdown spectroscopy using meta-learning

The sorting of coal and gangue is a critical process in coal production, playing a significant role in cost reduction, efficiency improvement, and green transformation of industrial workflows. To address the limitations of traditional supervised learning methods that are constrained by large-scale labeled data requirements, this study proposes a novel few-shot learning framework for coal and gangue classification based on laser-induced breakdown spectroscopy (LIBS) integrated with meta-learning techniques. Using NASA's Martian rock as source domain samples and coal and gangue as target domain samples, we proposed an enhanced meta-learning framework incorporating peak-aware and feature decoupling to highlight critical element responses and adapt to cross-domain classification tasks. The proposed method achieves an average classification accuracy of 90.08 % in 1-shot conditions and 95.04 % in 5-shot conditions, outperforming traditional machine learning methods by over 9 % and 4 %, respectively, and exceeding deep transfer learning approaches by 11 % and 2 %. Our method provides a core technological foundation for developing intelligent sorting equipment in mining operations.