A Novel Coal Dust Characteristic Extraction to Enable Particle Size Analysis

Abstract

Coal particle characteristic extraction is essential to prevent dust explosions. It is traditionally analyzed by manual screening, which is relatively time-consuming and cannot automatically capture outlining. To address this issue, an accurate and automatic particle image segmentation method is highly demanded in the smart mine. Thus, a novel characteristic learning approach that applied simplified VGGNet as a backbone network is investigated to learn the feature details of particle image sample. The sample set includes 3000 dust images captured from coal preparation plants. First, hierarchical features are extracted step by step on the improved VGGNet. Meanwhile, the feature maps obtained via convolution are sent to the dual attention mechanism module to determine the global feature weights, and the particle characteristic information is optimized. Afterward, an erosion-dilation module is applied to achieve dense texture separation in the deep feature map. Finally, the particles are segmented by upsampling on unpooling. The experimental results show that the proposed method achieves better precision, recall, and ${F}1$ with 0.8743, 0.8351, and 0.8543, respectively, than other previous methods. Compared with laser diffractometry, the maximum error $\varepsilon $ is 5.115% in the range of $R \le 75~\mu \text{m}$ , and this is consistent with the expected. The proposed method outperformed other state of the arts on segmentation results and can be applied in coal dust detection for enterprise as a viable alternative.