Advancing Coal and Gangue Classification: A Novel Approach Using 3d-Ct Data and Deep Learning

Abstract

Current coal gangue classification methods using deep learning rely mainly on 2D image data, such as optical and thermal imaging, which are limited to surface details and affected by external factors. To address this, we propose a 3D CT-based classification method that captures internal features, like mineral density and fractures, using a 3D deep learning framework. This approach enhances classification accuracy by replacing the MLP head with a Kolmogorov-Arnold Network (KAN). Comparative experiments with 2D CT and optical data show that 3D CT models exhibit faster convergence, better stability, and superior performance, achieving around 96% precision, recall, accuracy, and F1 score.

Keywords: 3D CT data, 3D CNN model, Coal and gangue classification, Kolmogorov Arnold Network, Deep learning