PCViT: A Pre-Convolutional ViT Coal Gangue Identification Method

Abstract

For the study of coal and gangue identification using near-infrared reflection spectroscopy, samples of anthracite coal and gangue with similar appearances were collected, and different dust concentrations (200 ug/m3, 500 ug/m3 and 800 ug/m3), detection distances (1.2 m, 1.5 m and 1.8 m) and mixing gangue rates (one-third coal, two-thirds coal, full coal) were collected in the laboratory by the reflection spectroscopy acquisition device and the gangue reflection spectral data. The spectral data were pre-processed using three methods, first-order differentiation, second-order differentiation and standard normal variable transformation, in order to enhance the absorption characteristics of the reflectance spectra and to eliminate the effects of changes in the experimental environment. The PCViT gangue identification model is established, and the disadvantages of the violent patch embedding of the ViT model are improved by using the stepwise convolution operation to extract features. Then, the interdependence of the features of the hyperspectral data is modeled by the self-attention module, and the learned features are optimized adaptively. The results of gangue recognition under nine working conditions show that the proposed recognition model can significantly improve the recognition accuracy, and this study can provide a reference value for gangue recognition using the near-infrared reflection spectra of gangue.

Keywords

• coal and gangue identification
• near-infrared reflection spectroscopy
• 1DCNN
• self-attention