Identification of Coal, Gangue, and Surrounding Rock Based on LIBS and Deep Learning
Zelin Yan, Dong Xiao
Abstract
As one of the most important energy sources in the world, efficient mining and application of coal is the basis for fully utilizing its effectiveness. During the mining stage, accurately identifying coal and surrounding rock is essential to prevent both oversegmentation and undersegmentation. During the beneficiation stage, separating gangue from coal is crucial to enhance coal utilization efficiency. Therefore, the accuracy of identifying coal, gangue, and surrounding rock has a direct impact on the yield and quality of coal. This study proposes a method for identifying coal, gangue, and surrounding rock using laser-induced breakdown spectroscopy (LIBS) and deep learning. The method initially utilizes clustering techniques to select key spectral bands and subsequently verifies their effectiveness. Then, the LIBS data are mapped into 2-D spectral images using the Gramian angular field (GAF). Innovatively, this article proposes a novel diffusion model to generate spectral images, which helps to overcome the difficulty of collecting labeled samples in mines. Finally, this article proposes an identification model, GAF-dimensional convolutional neural network (DCNN), which combines a multibranch structure with a cross-scale fusion strategy and extends the receptive field of the model using dilated convolution to efficiently learn fine-grained local spectral features and spatial-spectral features. The experimental results show that the proposed method achieves 92.44% in accuracy, precision, and recall metrics and keeps the test time consumption at a relatively low level. The method not only excels in accurately and efficiently identifying coal, gangue, and surrounding rock but also offers innovative approaches for enhancing coal utilization efficiency.